The central contention of this book is that the science of the mind has reached a point where multiple lines of empirical evidence, drawn from a wide variety of sources, converge to produce a resolution of the mind-body problem along lines sharply divergent from the current mainstream view.
The goal of the present chapter is to set the stage by sketching the evolution of mainstream psychology itself over the past hundred years, emphasizing modern developments and assessing critically where things presently stand.
The territory to be covered is vast and complex, so my account will necessarily be telegraphic and selective, and strongly colored by personal interests and experience—hence, “A View….” The center of perspective throughout is that of a working experimental psychologist, one whose professional experience includes psychology of language, human functional neuroimaging studies using both high-resolution EEG and fMRI methods, single-unit neurophysiological work in animals, and experimental parapsychology. I regret that I can claim only amateur-level acquaintance with relevant contemporary literature in philosophy of mind and language, although I recognize its value and unlike most of my fellow psychologists have made significant efforts to acquaint myself with it. The recent upsurge in efforts to bring these complementary perspectives simultaneously to bear on our central subject matter in a systematic and mutually informed way seems to me altogether welcome, and long overdue.2
The following three sections summarize the history of mainstream psychology in the English-speaking world from the advent of behaviorism to the present. So brief a sketch must necessarily be impressionistic, but I believe it is faithful to the main outlines of the subject as it has developed so far, and would be so regarded by most workers in the field. Useful general sources for readers wanting additional historical detail include Flanagan (1991), H. Gardner (1985), and Harnish (2002).
The history of scientific psychology in the 20th century can be characterized, somewhat cynically perhaps, as a movement toward progressively less unsatisfactory analyses of the mind. I will pass quickly over the first half of this history. Noam Chomsky remarked parenthetically during a lecture on linguistics in 1964 that in his opinion the first half century of American experimental psychology would end up as a footnote in the history of science. That was a characteristically provocative Chomsky remark, but even then it seemed more right than wrong. The historian of psychology Sigmund Koch, an early advocate of behaviorism who evolved into perhaps its most ferocious critic, has repeatedly derided the simplistic scientism of the period, and marveled at the degree to which behaviorism, having sprung into existence under the banner of a “consoling myth of natural science rigor and systematicity,” had so often “proceeded to liquidate its subject matter” (Koch & Leary, 1985, p. 942). During its heyday, from perhaps the 1920s to the late 1950s, behaviorism enjoyed extraordinary, almost monolithic, institutional and professional power. Even as late as 1990, the chairman of psychology at a major American university, in a symposium celebrating the centennial of the Principles of Psychology, characterized James’s great book as mainly illustrating what psychology should not be; its positive attributes consisted mostly of those few scattered passages where James’s observations corresponded to truths revealed by real psychology—the natural science devoted to analysis and control of behavior (Kimble, 1990).
Certainly one of the chief lessons to be derived here is that entire generations of industrious and able scientists can be captured by an ideology that is fundamentally unsound. I do not mean to say that behaviorism was all bad. Its central methodological impulse, emphasizing the importance of systematic empirical observation and measurement, was certainly healthy and remains so today. Even on that front, however, the early behaviorist program was unnecessarily narrow. Encouraged by the verificationist doctrine associated with the Vienna school of logical positivism (Ayer, 1952), behaviorists simply outlawed in principle all reference to anything not directly observable from the third-person standpoint. Stimuli, responses, and their supposedly lawful connections exhausted the scientifically legitimate subject matter. Even ignoring the often considerable difficulties in defining exactly what constitutes “a stimulus” or “a response,” however, this methodological asceticism was not warranted by any independently established conception of the nature of science. Indeed, philosophers of science soon abandoned verificationism in its narrowest construction, recognizing that even classical physics, the archetypal science, did not hesitate to postulate entities and processes that could not be observed directly, but only through their lawful connections to other things that could.
The experimental psychology that began to evolve after Watson’s 1913 manifesto can be viewed as a kind of operationalization of 19th-century associationist theories of the mind, in which “ideas” were replaced by behaviors, and complex behaviors were imagined as arising from simple ones through processes of conditioning and reinforcement. The bulk of this work was carried out with simpler organisms such as rats and pigeons, on the view that everything necessary for scientific psychology was present there, and in more accessible form. From the principles that emerged from such studies, it was hoped, we would eventually be able to build a psychology capable of accounting for all the complexities characteristic of human behavior. This specifically included what Descartes had insisted is uniquely ours, and a defining attribute of the mind—our use of language. A few mainstream behaviorists such as Edward Tolman and Egon Brunswick suggested that even the behavior of rats running their mazes might be guided by some sort of inner representation or map, but these suggestions were largely ignored.
Methodological behaviorism was subsequently reinforced for a time by a companion philosophical doctrine called logical or analytical behaviorism, which received perhaps its fullest expression in the influential book by Ryle (1949). It shared Watson’s objective of exorcising the mind, “the ghost in the machine,” but sought to achieve this objective by redefining men-talistic terms in terms of overt behavior, or dispositions to such behavior. Having a pain, for example, was to be construed as literally consisting in crying out, reaching for the aspirin, and so on. This relentlessly third-person approach to the mind seemed consistent with, and supportive of, the actual practices of behaviorist psychologists, but its problems as a philosophic doctrine soon became apparent.3 It proved extremely difficult in practice to specify, in finite detail and without covert reference to other mentalistic terms, the behavioral conditions in terms of which the original mentalistic terms were to be redefined. From a more commonsense point of view it also seemed to leave out precisely the things that are most important to us as human beings—in particular mental causation and our subjective conscious experience. We all know, for example, that pain and pain behavior simply are not the same thing. One can have a pain but not show it, or act as if in pain without actually being in pain. For these and other reasons analytical behaviorism fell generally out of favor, although echoes are still heard today—a primary source being Ryle’s student Daniel Dennett (1991).
Logical behaviorism gave way in the 1950s and 1960s to a family of positions known collectively as identity theory. Its basic doctrine is that the apparent correlation between mental states and brain states is to be interpreted in one particular way. Specifically, it holds that the relevant mental and physical states are in some sense identical, the same things viewed as it were from inside versus outside. This is regarded not as a logical necessity, but as a fact that we have discovered empirically, through advances in psychology and neuroscience, just as we have discovered that the morning star and the evening star are one and the same.
The identity doctrine came in two forms: The first and stronger form, formulated by writers such Herbert Feigl (1958), U. T. Place (1956), and J. J. C. Smart (1959), holds that mental states can be subdivided into discrete natural kinds or types, and that each of these types can be identified with a corresponding type of neural process. A stock example is the supposedly general relationship between “pain” and “excitation of c-fibers.” The weaker form of identity theory claims only that each individually occurring or token mental state is identical with some corresponding brain state. Note that type identity entails token identity, but not vice-versa.
Type-identity is a strong and interesting philosophic thesis which implies the possibility of reduction, and for these reasons many physicalists welcomed it; but it is certainly false. Quite apart from the difficulties of isolating appropriate “natural kinds” or types in either mental life or brain processes, it is certain that many such mental types, if they existed, would arise under wildly varying neurophysiological conditions. For example, linguistic behaviors involve mainly the left hemisphere in right-handed adults, but a mixture of both hemispheres or even mainly the right hemisphere in left-handers. The mind-brain system is in general enormously adaptable or “plastic.” For example, superior general intelligence and linguistic functioning have been observed in a man whose entire left hemisphere had been removed at age 5½ for control of seizures (A. Smith & Sugar, 1975). Fully functioning adults are also occasionally discovered who altogether lack major neural structures such as the corpus callosum or cerebellum, structures that are usually thought to be required for such functioning. In some well-studied cases of hydrocephalus, normal or even exceptional mental functioning has been found in persons who have only 3–5% of the normal volume of brain tissue (R. Lewin, 1980). And to take a still more extreme case, high-level forms of learning and memory are certainly present in the octopus, an invertebrate whose nervous organization is radically different from ours. It even lacks a hippocampus, the one structure that everyone agrees plays an essential role in mammalian memory systems (Chapter 4).
Type identity, therefore, has had little appeal for psychologists and neu-roscientists, who undoubtedly gravitate in the vast majority—to the extent they think about such things at all—to some sort of token-identity view. This was essentially the position advocated by James in The Principles, although he disavowed atomism in all forms and took pains to insist that the level at which the intersection is appropriately sought is that of whole momentary states of consciousness and whole momentary states of the brain. Token identity, on the other hand, has had relatively little appeal for philosophers. Among other things it creates a new and serious problem: If the same mental state, say a certain belief, can exist in combination with different sorts of physical states in brains, what is it about all those physical states that makes them “the same” as their common mental counterpart?
A philosophic response to this problem is the doctrine known as functionalism, first formulated by Hilary Putnam (1967).4 Having rejected type identity, on grounds that a given mental state might conceivably occur in extraterrestrial beings, or more relevantly in computers, Putnam went on to propose a novel solution to the problem just noted. His basic idea was to re-conceptualize mental states once again, this time not in terms of what they are made of, but in terms of what they do, their causal role in the functional economy of whatever sort of creature or entity is in question. Just as “cutting tools” can be implemented using rocks, metals, or laser beams, mental states are to be conceived as “multiply realizable”—that is, potentially instantiated in a variety of physical forms including not only token biological states of one or many brains, but also in computers and other suitable kinds of complex physical systems. On this view mental states become simply Xs, defined by their causal relations to stimuli, to other mental states, and to responses, and they can be identified as similar states to the extent they perform similar roles in their respective causal networks.
I will make only a few brief comments on this doctrine, which in various forms has dominated the philosophy of mind for almost 40 years. First, as originally formulated it was inherently and fundamentally third-person and behavioristic, albeit a refined behaviorism that admits the possibility of complex causal processes—mental causes, in effect—mediating between stimuli and responses. Like the earlier forms of behaviorism, it initially avoided all reference to consciousness and subjective features of mental life. This was widely felt to be unsatisfactory, however, and a large part of the subsequent history of functionalism consists of strained attempts to “naturalize” first-person phenomena of these sorts. Second, although functionalism readily affiliates itself with both physicalism in general and token identity theories in particular (and in most of its adherents probably still does), such afflliations are not an essential or inherent aspect of the doctrine. J. Fodor (1981a), for example, pointed out that functionalist principles might perfectly well apply to the operations of immaterial minds and the like, should any such things exist. Finally, it is fair to say, I think, that functionalism arose not so much sui generis in philosophy, but rather as a response to some exciting new developments which had already occurred within scientific psychology itself, and with which Putnam was certainly well acquainted. In any event, it was the confluence of these streams that defined the emergence of the 20th century’s most distinctive contribution to mind-brain theory, the “Computational Theory of the Mind” (henceforth, CTM). I turn now to the psychological dimensions of this story.
By the late 1950s discontent with behaviorism was rapidly spreading, as its inherent limitations became increasingly apparent. An influential paper by Lashley (1951) had exposed fundamental difficulties in the attempt to explain complex behavior, notably human linguistic behavior, in terms of linear chains of stimuli and responses. B. F. Skinner, the leading behaviorist, responded to this challenge, but his book on verbal behavior was subjected to a destructive and widely circulated review by Chomsky (1959). Most significantly of all, perhaps, a comprehensive state-of-the-science review organized by Sigmund Koch under the auspices of the American Psychological Association and the National Science Foundation resulted in a sweeping, 6000-page, six-volume indictment of the entire behaviorist platform (Koch, 1959–1963).
At the root of these discontents was a recognition that the old asso-ciationist explanatory principles, and their behavioral translations, were in principle unable to cope with the hierarchically organized and orderly character of human language and cognition. We needed a richer concept of mechanism. And as it happened, possible means of overcoming these limitations were just then becoming available, due to fundamental developments in the theory and practice of computation.
The old concept of a “machine”—and perhaps for most of us still the everyday concept—is that of a physical contraption which transforms energy by means of pushes and pulls involving gears, pulleys, shafts, levers, and so on. The fundamental insight underlying the modern developments is the recognition that these physical arrangements are really of secondary importance. The essential attribute of the machine is rather that its normal behavior is bound by rule. This insight opened the way to an enormous enrichment of our concept of mechanism, beginning with the contribution of logicians and mathematicians in the 1930s and 1940s and continuing into the present day. These developments, moreover, immediately began to have a profound impact on scientific psychology.
For example, it was quickly recognized that machines can transform data or “information” as well as energy, and that a machine can in principle utilize information about its performance to regulate its own behavior. These ideas had immediate and urgent practical application in the construction of servo-controlled antiaircraft systems during World War II, but possible theoretical implications for the understanding of behavior were also apparent. Rosenblueth, Wiener, and Bigelow (1943), for example, argued that from the point of view of an external observer a device constructed on this principle of “negative feedback” behaved purposively—that is, as a teleological mechanism. Thus, mechanism appeared to penetrate one of the last strongholds of old-fashioned vitalist thinking.5
These analogies were developed much more systematically by Wiener in his influential book Cybernetics, significantly subtitled Control and Communication in the Animal and the Machine. In addition to providing a general analytic theory of feedback control processes, Wiener (1961) provided numerous examples of physiological phenomena that seemed to fall within the province of the theory. Nevertheless, the direct applications of cybernetic theory at this level remained relatively limited. The real power of the ideas emerged later on, in conjunction with the extremely flexible applications technology provided by the digital computer.
To appreciate the full significance of these developments, it is necessary to follow the generalization of the concept of “machine” to its ultimate development in the hands of the British mathematician Alan Turing and several others. Turing devised an abstract representation that formalized his intuition of the core meaning of “mechanism,” as applied to the theory of computable functions. Any computation can be regarded as the transformation of a string of input symbols into a string of output symbols by a sequence of rule-governed steps. A procedure that is guaranteed to lead to the desired output in a finite sequence of steps is called an “algorithm” or “effective procedure.” Turing envisioned a machine consisting of a read/write head operating on an indefinitely extendable tape ruled off into squares. The behavior of the machine is completely specified by a set of five-part rules. Given the machine’s current state, and the input symbol written in the current square, these rules instruct the machine to change state, write a new symbol on the tape, and either remain where it is or move one square left or right. By altering the number of states, the size of the vocabulary, and the behavioral rules, an immense variety of behaviors can be realized by such devices. In fact, Turing argued persuasively that anything that would naturally qualify as an algorithm can be represented by a suitably constructed machine of this sort. He also proved rigorously that he could construct a “universal” Turing machine that would simulate the behavior of any other Turing machine. The intuitive notion of “effective procedure” was thus explicated in terms of the formal notion “realizable by a Turing machine.” That this is not an arbitrary result but in a fundamental sense exhausts the meaning of the concept of mechanism is strongly suggested by the fact that other workers such as Alonso Church and Emil Post arrived at provably equivalent results from widely different starting points.
Because of their utter simplicity, Turing machines do even very simple things such as adding two numbers in extremely cumbersome ways. Their significance is theoretical, not practical. But links to brain-mind theory were quickly forged. An influential paper by McCulloch and Pitts (1943) showed that networks constructed from idealized neurons could in principle implement logical behaviors of arbitrary complexity. Both they themselves and other workers further showed that equivalent capacities could be realized using richer elements that more nearly approximated the characteristics of real neurons. Thus it seemed likely that brains in principle have capacities equivalent to those of Turing machines. They might conceivably have additional capacities as well, but if so—and this is the essential point—these capacities may lie beyond the reach of understanding based on computational principles alone.
The final ingredient was provided by mathematician John von Neumann, who in 1947 invented the basic architecture of the modern stored-program digital computer. Von Neumann was entirely familiar with the theory of computability, and he was undoubtedly directly inspired as well by McCulloch, Pitts, and Wiener, all of whom participated with him in an important series of conferences on cybernetics sponsored by the Macy Foundation (Dupuy, 2000). It was evident to all that von Neumann’s new architecture provided in effect the logical capabilities of a universal Turing machine, but now at last in a practically useful form. To the extent that mind and brain are governed by formalizable rules, their activities could now in principle be modeled on suitably programmed general-purpose digital computers. To those sufficiently committed a priori to mechanistic principles, the very existence of any given class of behavior essentially entailed the possibility of such formalization. The relation of mind to brain could be conceptualized as analogous to the relation of computer software to computer hardware, and the mind-brain problem would simply disappear.
There were other more specific theoretical results that further strengthened this emerging point of view. Consider, for example, some of the early results in theoretical linguistics. Chomsky (1963) and others showed that the possible classes of formal models of language (generative grammars) formed a hierarchy, in which the weakest or most highly constrained class (finite-state grammars) was obtained from the strongest or least constrained class (unrestricted rewriting systems) by the application of progressively more severe constraints on the form of the permissible rules of the grammar. The resulting hierarchy of grammars, it turns out, also corresponds to a hierarchy of classes of automata derived from Turing machines in parallel fashion. Formal results from automata studies thus transferred to the analysis of candidate grammatical theories. Chomsky (1957) was able to show that then-existing psychological and linguistic proposals for theories of language, when formalized, corresponded to the weakest or most constrained members of the hierarchy of grammars, and that these grammars were in principle too weak to account for systematic structural properties of many kinds of sentences found in natural languages such as English. He was thus led to his famous theory of transformational grammar as the weakest class of theory which is still strong enough to account for the known grammatical facts of language. The result that the corresponding automata are weaker than Turing machines greatly strengthened the presumption that linguistic behavior might be formalizable for computer modeling.
The central idea that minds, brains, and computers could fruitfully be regarded as variants of a more general class of information-processing mechanisms quickly took root, even among neuroscientists (W. J. Freeman, 1998; von Neumann, 1958). The ground was very well prepared. Indeed, these developments seem to me an inevitable outcome of our Western scientific tradition. This is not meant disparagingly, however. I have stressed these results about Turing machines and so on precisely to underscore the impressive depth of the theoretical foundation on which all the ensuing developments rest, a foundation which I feel has not been adequately appreciated by many critics of this kind of work, such as Edelman and Tononi (2000), nor even by some of its enthusiastic supporters, such as H. Gardner (1985).
In practice, the applications came a bit slowly at first. In part this was due to purely technical factors. The early computers were small, slow, and highly prone to malfunction. More importantly, in the early days programming a computer was an exasperating business requiring detailed familiarity with low-level details of the hardware organization. Indeed, the basic elements of the available programming languages referred to data structures and operations virtually at the hardware level. As the technology advanced, however, machines grew larger, faster, and more reliable, and so-called “higher-order” languages were created, such as FORTRAN, whose primitives referred to data structures and operations at a level relatively natural for human problem-solvers. Special-purpose applications programs written in a higher-order language congenial to the user could then be translated by general-purpose “compiler” or “interpreter” programs into the internal language of the computer for execution by the hardware.
I mention these details because they relate to the other main reason for the delay, which is more theoretical in nature and involves a basic question of strategy. The fantastic complexities of the brain can obviously be studied at many different levels from the cellular or even sub-cellular on up. At what level shall we seek scientific explanations of human mental activity? Many scientists, particularly those working at lower levels of the hierarchy of approaches, assume that events at the higher levels are in principle reducible to events at lower levels, and that reductive explanations employing the concepts of the lower levels are necessarily superior or more fundamental. Like many other psychologists, I strongly disagree with this view.
Consider, for example, the problem of understanding the behavior of a computer playing chess under the control of a stored program. It seems obvious that we might observe the behavior of the computer’s flip-flops forever without gaining the slightest real understanding of the principles that are embodied in the program and explain its behavior. One of the essential characteristics of both human and animal behavior is that functional invari-ance at a higher level may be coupled with extreme diversity at lower levels. For example, the rats whose cortex Lashley (1950) progressively destroyed in efforts to locate the memory trace could wobble, roll, or swim to their food box, and I can write a given message with either hand, or probably even with my feet if necessary. Attempted explanations based on activities of the participating muscle-groups, neurons, and so on would never get to the essential common feature, which is the approximate invariance of the final behavioral outcome.
Thus it seems appropriate in general to seek a higher and perhaps distinctively “psychological” level of explanation for human cognitive processes. For the computer simulation approach this involves identifying an appropriate set of elementary information structures and processes that seem powerful enough in principle to account for the relevant behaviors. The hypothesized structures and processes should perhaps also be potentially realizable in the brain, given its known properties—or at least not demonstrably inconsistent with these properties—but successful use of the computer as a tool of psychological understanding does not require the obviously false presumption of literal identity between digital computers and brains.
By the late 1950s and early 1960s, a number of higher-order programming languages had been created which emphasized the capacities of computers as general-purpose symbol-manipulating or information-transforming devices, rather than their more familiar “number-crunching” capacities. These languages (such as IPL-V, LISP, and SNOBOL) provided facilities, for example, for creating and manipulating complex tree-like data structures consisting of hierarchically ordered and cross-referenced lists of symbols. Structures of this sort played a central role in theoretical linguistics, and in this and other ways the new languages seemed to many workers to fall at about the right level of abstraction to support realistic efforts to model important aspects of human cognition.
Previous generations of workers had been obliged either to try to force human mental processes to conform to artificially simple but relatively rigorous behavioristic models, or to lapse into the uncontrolled introspection and mentalistic speculations of an earlier era. Now suddenly we were provided with a conceptual and technical apparatus sufficiently rich to express much of the necessary complexity without loss of rigor. The “black box” could be stuffed at will with whatever mechanisms seemed necessary to account for a given behavior. A complicated theory could be empirically tested by implementing it in the form of a computer program and verifying its ability to generate the behavior, or to simulate a record of the behavior. Progress toward machine intelligence could be assessed by means of “Turing’s test”—the capacity of a computer program to mimic human behavior sufficiently well that a physically remote human interlocutor is unable to distinguish the program from another human in an “imitation game” (Turing, 1950). The seminal modeling ideas of Craik (1943) could at last be put into practice.
Enthusiasm for the computational approach to human cognition fairly crackles through the pages of influential early books such as G. A. Miller, Galanter, and Pribram (1960), Lindsay and Norman (1972), and Newell and Simon (1972). Their enthusiasm seemed justified by the ensuing flood of theoretical and experimental work based on these ideas. In addition to the specific efforts at computer modeling of cognitive functions that is my main concern here, the rise of the computer-based information-processing paradigm also stimulated a healthy reawakening of broader interests in many of the old central concerns of psychology, such as mental imagery, thinking, and even consciousness. I must also acknowledge here that I myself initially embraced the computational theory practically without reservation. It certainly seemed an enormous step forward at the time. Fellow graduate students likely remember my oft-repeated attempts to assure them that the CTM would soon solve this or that fundamental problem in psychology. But all was not well.
With the appearance of suitable higher-order languages, numerous research groups set to work to endow computers with capacities for various kinds of skilled performance, including game-playing (especially complex games such as chess), problem-solving (for example, proving theorems in the propositional calculus), pattern recognition (such as recognizing sloppy hand-written characters), question-answering (in restricted domains such as baseball), and natural language translation. An important strategic difference quickly appeared, dividing these efforts roughly into two streams often called computer simulation (CS) and artificial intelligence (AI), respectively. Workers in CS remained faithful to the aim of increasing psychological understanding, in that they sought not only to reproduce particular kinds of human performance, but also to identify testable models of the means by which humans achieve that performance. AI workers, by contrast, disavowed any direct interest in psychology and sought rather to achieve high-level performance by whatever means possible.
I will not attempt to review the substantive accomplishments here.6 Suffice it to say that many of the individual contributions represented considerable intellectual and technical achievements. Especially dazzling for me was “SHRDLU,” a virtual robot that could interact with its handler in more or less ordinary-appearing language while carrying out complex sequences of requested operations on toy blocks of varied color, size, and shape (Wino-grad, 1972).
However, without intending to disparage these attainments I must add that they still fell very far short of what anyone could plausibly describe as general intelligence. I mention this only because of the extremely inflated image which many outside observers had at the time of the progress of this work, an image aggressively promoted by some of the research workers themselves. Many extraordinarily grandiose statements and predictions were being bandied about on the basis of very modest concrete accomplishments. Of course the predictions could conceivably have been correct; after all, the theoretical foundation is deep, and the work was still in its infancy. Perhaps we were simply in the early stages of an evolutionary process in which machines would inevitably attain at least the equivalent of human cognitive abilities.
My own confidence in the new paradigm was severely shaken, however, by extensive and sobering experience associated with my dissertation research in the area of natural language processing. The group with which I was working was principally concerned with the applied technical problem of reducing lexical ambiguity in English text, the practical aim being increased precision and power of automated computer-based content analysis procedures. In approaching this problem, we constructed an alphabetized concordance of some half-million words of “typical” behavioral science texts sampled from a variety of sources. This keyword-in-context listing (KWIC) identified the most frequently occurring words and supplied information about their typical patterns of usage. It turned out that about 2,000 types (in this context, dictionary entries) covered around 90% of the tokens (occurrent words) in an average running text. The striking regularities of word usage evident in the KWIC enabled us to build for each such dictionary entry a small computer routine that could scan the context in which each new token instance of that entry appeared and attempt to determine which member of a pre-established set of meanings was actually present. Crude as it was, the resulting system worked surprisingly well, achieving about 90% accuracy in a large new sample of text (E. F. Kelly & Stone, 1975). So at a practical level the project was rather successful.
My own primary interest, however, lay in determining whether the brute facts of everyday language as we were seeing them could successfully be captured by existing computationalist theories of semantic representation. The main such theory at the time was that of J. J. Katz and Fodor (1964). Their theory built upon the central notions of Chomsky’s transformational linguistics and embodied the essential doctrines of the classical computationalist program as subsequently formalized by J. Fodor (1975), Pylyshyn (1984), and others. Specifically, it depicted determination of the meaning of a sentence as resulting from a rule-bound calculation, governed by the syntactic analysis of that sentence, which operates upon pre-established representations of the meanings of its constituent words. The possible meanings of the words themselves were treated as exhaustively specifiable in advance, with their essential semantic contents represented in terms of an underlying universal set of discrete “semantic markers”—in effect, atoms of meaning—or logical structures built out of such markers. In a nutshell, syntax was to provide for the generativity of language, our ability to make “infinite use of finite means,” and semantics would go along for the ride.
Although Katz and Fodor had made their account appear at least semi-plausible for a few carefully contrived examples, it failed totally to account for obvious and pervasive features of the language in our database, language as actually used by ordinary speakers for ordinary purposes. For one thing, it was clear that resolution of lexical ambiguity routinely relied upon all sorts of linguistic and non-linguistic knowledge, and not just on some supposedly “essential” knowledge of pre-established word meanings. More importantly, sentence interpretation could not plausibly be viewed as a matter of context-driven selection from large numbers of pre-established and highly specific word meanings. Rather, it clearly involved generation of appropriate interpretations based on much smaller numbers of more general meanings (see, e.g., the entries provided by different dictionaries for common workhorse words such as among, find, fine, and line, and note that all major parts of speech display these properties). Any scheme based on atomization of meaning would necessarily fail to capture what to me had become the most characteristic property of word-meaning, a felt Gestalt quality or wholeness, at a level of generality that naturally supports extensions of usage into an indefinite variety—indeed whole families—of novel but appropriate contexts. The existing proposals could only represent the content of a general term such as “line” by some sample of its possible par-ticularizations, and in so doing rendered themselves systematically unable to distinguish between metaphorical truth and literal falsehood. It seemed especially ironic that Katz and Fodor, for all their transformationalist invective about the need to respect the productivity of language, had erected a semantic theory which in effect elevated the commonplace to a standard of propriety and denied creativity at the level of words or concepts. Their theory failed to account even for identification, let alone representation, of linguistic meaning (E. F. Kelly, 1975).
I also noted with a certain degree of alarm that this crucial property of generality underlying the normal use of words seemed continuous with de-velopmentally earlier achievements. Skilled motor acts and perceptual recognition, for example, require similar on-the-fly adaptations of past learning to present circumstance. It seemed at least vaguely conceivable that these difficult but undeniably fundamental qualities of embodied action and cognition might somehow be rooted in lower-level properties of the nervous system as a particular kind of computing machine. Early discussions had emphasized similarities between brains and digital computers, for example treating the all-or-nothing neural spike discharge or action potential as the equivalent of a digital relay. In more recent times, however, we had become increasingly sensitized to the ubiquitous presence in real nervous systems of inherently more continuous or “analog” processes, such as the spatial and temporal summation of neural input that leads to spike formation, and the rate and pattern of the resulting spike discharges. Workers in CS and AI had been relying heavily on their doctrine of “multiple realizability,” assuming perhaps too cavalierly that they could safely disregard such low-level “hardware” details and pitch their efforts at a level of abstraction that happened to be congenial both to them and to the available computers. What seemed now to be emerging, however, was that these low-level neurophysiological properties might enter into the overall computational possibilities of the brain in a much more fundamental way than previously suspected. Perhaps the “missing” cognitive attributes were dependent upon such low-level characteristics. Most neuroscientists, of course—even while adopting more or less wholesale the new computationalist or “information-processing” paradigm—had always presumed this would be so.
The net result of all this was to make me much more skeptical and pessimistic regarding the prospects for computational cognitivism as it was then being practiced. To be sure, others had noted similar problems and were trying to do something about them. The importance of incorporating more general knowledge of the world into language-processing models, for example, had already begun to be recognized, and new formal devices were being introduced to represent what the computer needed to know (what we ourselves know) about various sorts of “typical” situations it might encounter. But it seemed clear to me that all of these knowledge-representation devices, such as “frames” (Minsky, 1975), “scripts” (Schank & Colby, 1973), and “schemata” (Neisser, 1976), suffered essentially the same problems I had identified in the Katz and Fodor account of word meaning. Specifically, they required the possible scenarios of application to be spelled out in advance, in great but necessarily incomplete detail, and as a result ended up being “brittle,” intolerant of even minor departures from the pre-programmed expectations. I also pointed out that Winograd’s undeniable success with SHRDLU had derived in considerable part precisely from his careful a priori restriction of the problem domain in a manner that allowed him essentially to circumvent this whole complex of very difficult issues.7
Many of the themes just sounded have been confirmed and amplified in more recent work. On the positive side, our knowledge of the content, organization, and development of the human conceptual system has increased enormously. The old Socratic idea that concepts must be defined in terms of necessary and sufficient features has given way to a recognition of the role of perceptual-level examples, prototypes, and family resemblances in the content of real human concepts (Medin & Heit, 1999; Rosch & Lloyd, 1978; E. E. Smith & Medin, 1981). The contributions of a fundamental human capacity for metaphorizing at levels ranging from everyday language to the highest flights of creativity, are also now more widely appreciated (Gentner, Holyoak, & Kokinov, 2001; Hofstadter & FARG, 1995; Holyoak & Tha-gard, 1995; Lakoff, 1987, 1995; see also our Chapter 7). Computer language-processing systems have tended to move, as I predicted, toward a relatively simplified and “surfacy” syntax coupled with richer representations of the lexicon, and they sometimes attempt to resolve residual ambiguities with the aid of statistical data derived from large databases.
More crucial to the concerns of this chapter, however, are major subsequent events on the negative side. Shortly after I completed my dissertation, Hubert Dreyfus (1972) systematically questioned both the progress and the prospects of CS and AI. He began by reviewing the early work in game-playing, problem-solving, language translation, and pattern recognition. Work in each domain was characterized by a common pattern consisting of encouraging early success followed by steadily diminishing returns. Subsequent work in artificial intelligence, according to Dreyfus, had fared little better, achieving its limited successes only by operating in artificially constrained environments that do not exemplify fundamental difficulties handled with conspicuous ease by everyday human intelligence. Dreyfus argued that the extreme efficiency of human intelligence—which becomes progressively more apparent as we move toward problem domains more typical of its ordinary application—rests on a complex of interrelated abilities which he termed fringe consciousness, ambiguity tolerance based on efficient use of context, essential-inessential distinctions, and perspicuous grouping. Most critically of all, human cognitive performance is characteristically guided by insight, an overall grasp of the problem situation, with essential aspects at the foreground of attention set against an organizing but largely implicit background. Phenomenologically, the situation is primary; specific facts or features of the situation may only become evident through a deliberate attentive effort of a quite secondary sort. By contrast, Dreyfus argued, for the computer all facts must be specified in advance as explicit bits of atomic data; whatever crude representation of the situation the computer can achieve is necessarily constructed by explicit calculation upon these situation-independent facts. Ironically, AI seemed to have adopted the conceptual framework of Wittgenstein’s Tractatus shortly after the realities of language use had driven Wittgenstein himself to abandon it. Building upon these and related observations, Dreyfus attempted to establish his main thesis, that central human mental skills are in principle not reproducible on digital computers.
Dreyfus’s concerns clearly overlapped with mine, and I therefore had no doubt that he had identified a cluster of problems which at the very least constitute extremely difficult problems of practice, although I was not yet convinced that these problems were insoluble in principle. Hardcore AI partisans experienced no such unease, attacking Dreyfus immediately and viciously as merely an ignorant outsider. Shortly afterward, however, Joseph Weizenbaum, a prominent AI insider,8 expressed broadly similar kinds of misgivings, although he was less confident than Dreyfus in theoretical arguments against the possibility of machine intelligence, and relied primarily on a moral argument to the effect that there are many kinds of things, including psychotherapy, which computers should never be permitted to perform (Weizenbaum, 1976).
The most significant by far of these rumblings from within, however, occurred a decade later when a consummate insider, Terry Winograd himself, publicly defected from the program of classical AI. Winograd and Flores (1986) explicitly embraced most of the points already raised above, emphasizing in particular that large parts of human mental life cannot be reduced to explicit rules and therefore cannot be formalized for production by a computer program. Their detailed and fully-informed argument should be required reading for anyone interested in these issues.9 The best we can hope for along classical lines, they concluded, is special-purpose expert systems, adapted to carefully circumscribed domains that lend themselves to such formalization. Current examples, of course, include things like chess-playing, integration and differentiation of mathematical formulae, and perhaps certain areas of medical diagnosis.
Since the late 1970s psychology has taken a strongly biological turn, and cognitive science has evolved into cognitive neuroscience. This came as somewhat of a surprise to persons like myself who had been reared in the intellectual tradition of classical or “symbolic” cognitivism and had expected to do all their proper business with little or no reference to biology. When I was a graduate student, for example, the required course in physiological psychology consisted essentially of a smattering of neuroanatomy followed by a lot of boring stuff about appetitive behavior in lower organisms, and that was pretty much that.
I will single out four main threads of this evolution. The first arose within classical cognitivism itself. As indicated above, there was a lot of ferment in the early days as researchers sought to identify appropriate forms of representation for the sorts of knowledge we humans can bring to bear in our thinking, speaking, and so on. Largely reflecting its own historical origins and the available means of representation, cognitive theory initially focused almost exclusively on linguistic or propositional forms of knowledge representation. Mental imagery, however, had recently been readmitted onto the roster of acceptable research topics (R. R. Holt, 1964), and a number of important new experimental studies were being carried out, especially by Roger Shepard, Stephen Kosslyn, and various others. On the strength of these investigations Kosslyn (1973) ventured to advance an information-processing theory of visual imagery in which the underlying knowledge structures hypothesized to support the generation of imagery were themselves essentially pictorial and spatial in nature. This provoked an intense reaction from Zenon Pylyshyn (1973), who reaffirmed the classical cognitivist view that all knowledge is linguistic or propositional in character (J. Fodor, 1975). The ensuing imagery debate has raged on more or less ever since, its intricate experimental details of interest primarily to the participants.10 For present purposes the critical event was an important theoretical paper by John R. Anderson (1978). Anderson argued that since both kinds of representations (and potentially many others as well) could be made internally coherent, and would lead to identical behavioral predictions, a fundamental theoretical indeterminacy had emerged. Considerations of parsimony and efficiency might lead us to prefer one such theory to its competitors, but only physiological observations could potentially determine which was in fact correct. Thus it became evident, more generally, that neurophysiological data can sometimes provide important constraints on psychological theory.
The second thread was the emergence and consolidation following World War II of neuropsychology as a scientific discipline. An outgrowth and companion of the medical discipline of neurology, neuropsychology seeks to gain insight into the nature and operations of the mind through careful observation and analysis of the mental disturbances, sometimes highly specific and bizarre, that are produced by gunshot or shrapnel wounds, strokes, degenerative diseases, and other forms of injury to the brain.11 The general thrust of this work was to suggest that skilled cognitive performances of all kinds characteristically involve cooperation of a number of localized cortical and subcortical regions of the brain, each presumptively specialized for some particular role in the overall performance. Classical cognitivism quickly adapted to this emerging picture of things, assimilating its fundamental theme of the mind as a computational or information-processing mechanism to a “modular” view of its components and internal organization (J. Fodor, 1983; Pinker, 1997).
The third development was the advent and maturation of the functional neuroimaging technologies mentioned in the Introduction, which enable us to observe directly, without opening up the skull, the activity of the working human brain. Two principal classes of such methods are currently available, although others are under development.12 The first provides measures of the electric and magnetic fields directly generated by neuroelectric activity in populations of suitably located and oriented cortical neurons (electroencephalography [EEG]; magnetoencephalography [MEG]). The second provides measures of hemodynamic or metabolic consequences of neural activity such as local changes of blood flow, glucose consumption, or blood oxygenation (especially positron emission tomography [PET]; and functional magnetic resonance imaging [fMRI]). All these technologies are complex and expensive, and they tend within and between classes to have complementary strengths and weaknesses (see, e.g., Nunez & Silberstein, 2000; Toga & Mazziotta, 1996; Wikswo, Gevins, & Williamson, 1993). Together, they undoubtedly constitute a major methodological advance for cognitive neuroscience. Indeed, scarcely an issue now goes by of any cognitive neuroscience journal that does not contain one or more papers featuring images outfitted with colored spots identifying regions of “significant” brain “activation” produced by some stimulus or task.13
The final thread is the one most directly germane to our primary subject, the computational theory of the mind (CTM). Specifically, discouragement with the progress of classical or symbolic cognitivism led in the 1980s to an enormous resurgence of interest in a fundamentally different style of computation that also seems more directly comparable to what actually goes on in brains—namely, propagation of activity through large networks of basically similar elementary units.
I say “resurgence” because this approach actually harks all the way back to the seminal work of McCulloch and Pitts (1943), Hebb (1949), and others. There had been some promising early steps in this direction, notably the work of Rosenblatt (1962) on “perceptrons,” but mainstream interest had largely faded under the impact of a devastating critique by AI partisans Marvin Minsky and Seymour Papert (1968). These authors, mathematicians by training, proved rigorously that simple one-layer perceptrons cannot compute even simple things such as the elementary logic function “exclusive OR” (that is, either A or B but not both). In their conclusion, summing up the implications of these results, they conjectured that more complicated networks would fare little better. In effect they urged the field to stay focused on the CTM in its classic symbol-processing form. One of the great ironies here, as pointed out by Dupuy (2000), is that whereas John von Neumann, the great mathematician, had been interested in expanding the possibilities of computation by reference to the actual behavior of the brain, Warren McCulloch, a neuroscientist, had in effect abstracted from the brain a logic machine; and now here was Minsky, who had been von Neumann’s doctoral student at Princeton, promoting McCulloch-style symbolic computation in preference to the neurophysiologically grounded approach advocated by his own illustrious mentor.
At any rate the net effect was to drive neural-network models temporarily to the margins of the field, although a few dedicated souls such as James Anderson and Stephen Grossberg soldiered on. The subject burst back into the mainstream, however, with the publication of a two-volume handbook on parallel distributed processing (PDP) or “connectionism” as it came to be called (Rumelhart & McClelland, 1986).14 Connectionism has subsequently emerged as a serious rival to classical cognitivism, but I will give here only the briefest account of these developments.
The fundamental faith of connectionists is that intelligence emerges from the interactions of large numbers of simple processing units organized in a network of appropriate structure. By modifying features of network architecture such as the number of elementary units, the number of layers, their connectivity patterns (feed forward only vs. recurrent), the rules for activating units (simple thresholds, sigmoid functions), and the rules for modifying the connection strengths among units in light of network performance or experience (Hebb, delta, generalized delta), an enormous variety of interesting behaviors can be produced. Networks have proved especially good at some things at which classical models were conspicuously bad, such as pattern recognition, perceptual learning and generalization, and filling in of incomplete input. They also display psychologically interesting and desirable properties such as content-addressable memory and “graceful degradation”—the tendency, as in Lashley’s rats, for performance to decline smoothly and continuously as units are progressively removed. It is perhaps not surprising, given these properties, that unbridled optimism soon reappeared within the field. One leading connectionist, Smolensky (1988), has proclaimed that “it is likely that connectionist models will offer the most significant progress of the past several millennia on the mind/body problem” (p. 3). Many contemporary psychologists agree, and even some philosophers of mind, including in particular Daniel Dennett and the Church-lands, are no less enthusiastic.
Significant problems have also come to light, however. Although network models are often said to be “neurally inspired,” the current level of neurophysiological realism is typically very low. Both the “neurons” themselves and their connectivity patterns are routinely idealized and distorted, and the most successful learning rule (“back propagation” or generalized delta) still has no generally recognized counterpart in the nervous system (Crick, 1994). Models often have large numbers of free parameters which must be adjusted to fit specific situations, raising doubts about their generality. Similarly, generation of a targeted behavior is sometimes strongly and unrealistically dependent on the exact content and order of previous network experience or training. Many observers, including Papert (1988), also worry that problems of catastrophic interference between previously learned behaviors will emerge as networks are scaled up to more realistic dimensions. Most important of all, networks have particular difficulty capturing precisely those characteristics of human cognition to which the classical cognitivist approach seems best suited—features such as its hierarchical organization, temporal orderliness, and productivity. Indeed, despite some new technical wrinkles, connectionism can be viewed as amounting in large part to a modern revival of 19th-century associationism. To the extent that this characterization is correct, connectionism therefore inherits the very same problems that led to the rise of classic cognitivism in the first place (J. Fodor, 2001; Pinker & Mehler, 1988).
The last 20 years or so of computational modeling of the mind have been dominated by the attempts of these two warring paradigms to refine themselves and work out their relationship. Classical symbol-manipulation modelers, for example, are trying to find ways to make their models more adaptable and less brittle, while connectionists are looking for better ways of generating orderly behavior in the absence of explicit rules. Each paradigm clearly aspires to absorbing the other, but some workers are also exploring “hybrid” computational systems in which the connectionist part takes care of things like perceptual learning and category formation—the subsymbolic level or “microstructure” of cognition—while a more classical symbol-manipulation part deals with things like problem-solving and language. It is not at all clear how things will sort out, although the connectionist faction seems presently ascendant.
I must also mention, however, one further recent development emerging from within connectionism itself, one that threatens in effect to devour it from the other side. This movement, dynamic systems theory, has close ties to modern technical developments in the physics and mathematics of complex systems. There can be no doubt that the brain, with its vast network of reciprocally interconnected and non-linear elements, is an example of the kind of systems to which these new developments apply. Dynamic systems theorists often use typical connectionist technical apparatus to synthesize network models, but they bring to bear additional mathematical and graphical tools to characterize and analyze the temporal dynamics of network behavior. The most distinctive property of this school, however, is its more radical theoretical position. In contrast with mainstream connectionism, which it regards as a kind of halfway-house or unholy compromise with classical cognitivism, it advocates abandoning altogether cognitivism’s next deepest commitment after computation itself—namely, the concept of a representational level From the dynamicist point of view what goes on inside our heads is nothing like what the classical formulation presumes, with computational processes operating on stored symbolic knowledge-representations of one or more sorts. There is instead only the mechanical operation of a vast neural network in accordance with deterministic rules. As Dupuy (2000) observed, this is clearly the direction in which the early cyberneticians such as von Neumann were headed. It is also the view of leading contemporary neuroscientists and modelers such as Edelman and Tononi (2000) and Walter Freeman (1999, 2000).15
I shall not attempt here to adjudicate among these various schools of computationalism on their own terms. Readers interested in getting a more complete picture of this enormous and thriving industry in all its facets can probably do no better than to sample the contents of the MIT Encyclopedia of the Cognitive Sciences (MITECS) (R. A. Wilson & Keil, 1999). For the record, my own overall sense is that the accomplishments, although certainly substantial, continue to lag the “hype” by a wide margin. This seems especially true in the area of language processing, which, as Descartes and Chomsky in particular have insisted, should be viewed as a crucial test case. Despite my skepticism, however, I believe that persons of all philosophic or methodological persuasions should strongly support further efforts along all of these lines. As the computational theories evolve, they are constantly sharpening our understanding not only of what the various kinds of computational models can do, but also of what we ourselves do in a broad range of situations. Even apart from its ultimate success or failure, each new attempt to extend computer modeling into a novel domain of human cognitive performance forces us in the first instance to try to understand in detail what capacities are presupposed by that performance. As our understanding of the capacities of minds deepens and becomes more detailed, it should become increasingly possible to judge precisely which of those capacities may be explainable, and to what extent, in terms of such models.
There are good reasons, however, to believe that computationalism alone, in any of its forms, cannot provide a complete account of the mind. I turn next to these. All readers, but especially newcomers to these issues, are advised that from this point forward my discussion becomes more tendentious and increasingly at odds with current mainstream opinion.
Many persons both inside and outside of psychology have expressed reservations of one sort or another about computationalism, but without doubt the most sweeping, searching, and sustained attack on the CTM has come from Berkeley philosopher John Searle.
Briefly, Searle has two principal arguments.16 The first, the famous “Chinese Room” argument, was directed at the central claim of strong artificial intelligence—specifically, the idea that running a computer program can of itself be sufficient for, or constitutive of, understanding (Searle, 1980, 1984, 1990). The essence of this thought experiment is that a person who knows no Chinese, but who appropriately answers questions in Chinese by virtue of manipulating symbols according to the rules of a (hypothetical) computer program, would not thereby understand any part of the resulting “conversation.” Computing is by definition purely syntactical, and consists only of manipulation of uninterpeted formal symbols according to the explicit rules of a program. The occupant of the Chinese room does all that but understands nothing. Hence, semantics is something over and above syntax, and strong AI is false.17 Moreover, the Turing test is clearly inadequate; simulation is not duplication, and to act as if one understands Chinese does not guarantee that one does.
Classical cognitivism in its more general forms, however, might seem to escape the force of this simple argument, insofar as the computational component is no longer regarded as constituting or providing the semantics, but only as manipulating, in a manner that mirrors and “respects” their content, semantic representations that are provided elsewhere in the theory (as, e.g., in the theory of J. J. Katz & Fodor, 1964). Against this move Searle mounts a further and deeper argument. The most basic underlying assumption of classical cognitivism is that the brain, although certainly different from von Neumann machines, literally is a computer. However, Searle argues, “computation” and “information-processing” are not observer-independent features of the world that we empirically discover, like mass, gravity, photosynthesis, and digestion. Rather, these are observer-relative properties that we assign to certain systems for our own purposes. In short, the claim that the brain is a computer is not false, but incoherent. Not only is semantics not intrinsic to syntax, as shown by the Chinese Room, but syntax is not intrinsic to physics, or to the neurophysiology of the brain (Searle, 1992, 1997).
The details of Searle’s arguments, and the firestorms of controversy they have provoked over the years, are far too numerous and complex to sort out here, but I wish to make a few summary remarks.
First, like other critics of the CTM such as Dreyfus (1972) and Penrose (1989, 1994), Searle has been startled by the intensity of the reactions he provokes:
Oddly enough I have encountered more passion from adherents of the computational theory of the mind than from adherents of traditional religious doctrines of the soul. Some computationalists invest an almost religious intensity into their faith that our deepest problems about the mind will have a computational solution. Many people apparently believe that somehow or other, unless we are proven to be computers, something terribly important will be lost.18 (Searle, 1997, p. 189)
Second, this “passion” undoubtedly accounts for the apparent inability of many psychological and philosophical defenders of the CTM even to report Searle’s message accurately, let alone to appreciate its force. Without dwelling upon the details, see for example Bechtel and Abrahamsen (2002, pp. 303–304), Boden (1990), Chalmers (1996, chap. 9), H. Gardner (1985, pp. 171–177), Harnish (2002, chap. 9 & chap. 13), and Pinker (1997, pp. 93–96). Many others, such as Minsky (1985), Newell (1990), and Thagard (1998), essentially ignore him altogether. Searle defends himself much better against his critics than I could do on his behalf, and fortunately he has taken a number of opportunities to do so. In addition to the original debate over the Chinese room (Searle, 1980), see in particular his encounters with the Churchlands (Searle, 1990) and with Dennett (Searle, 1997).
Third, although the initial targets of Searle’s arguments were strong AI and cognitivism in its now classic “symbol-processing” formulation, there is no doubt that his arguments apply to computationalism in all of its forms. Searle made this explicit for standard forms of connectionism in his exchange with Dennett (Searle, 1997), and I think he would certainly do so for dynamic systems theory as well, and for the same reasons. The fundamental commonality of the different sorts of models is underscored by the fact that they all are typically implemented in the same sort of physical architecture, the standard von Neumann architecture used in garden-variety personal computers. Specifics of the computing architecture can profoundly influence the speed and efficiency of computation, but this does not alter its fundamental character and has no bearing on the core theoretical issues.19
In sum, I feel sure that Searle’s negative judgment regarding the computational theory of the mind will be sustained. Like Nagel (1993a), I have come to regard the CTM as one of the dominant illusions of our age. I hasten to add that this judgment has little practical import for cognitive science; for as Searle himself points out, and as I remarked above, use of computers to model particular cognitive functions is a perfectly legitimate and undoubtedly useful scientific activity. But for a fundamental theory of the mind we must look elsewhere.
To the extent that physicalist accounts of the mind are identifiable with the CTM, Searle and his allies have already discredited them. However, physicalism has not yet exhausted its resources. Searle’s further and positive contribution is to advocate a novel philosophic position which he calls “biological naturalism,” a position which in fact closely approximates the views of many leading neuroscientists. I turn next to this.
To set the stage I will quote the most trenchant statement known to me regarding the current status of the mind-body problem. It is from philosopher Thomas Nagel (1993b):
The mind-body problem exists because we naturally want to include the mental life of conscious organisms in a comprehensive scientific understanding of the world. On the one hand it seems obvious that everything that happens in the mind depends on, or is, something that happens in the brain. On the other hand the defining features of mental states and events, features like their intentionality, their subjectivity and their conscious experiential quality, seem not to be comprehensible simply in terms of the physical operation of the organism. This is not just because we have not yet accumulated enough empirical information: the problem is theoretical. We cannot at present imagine an explanation of colour perception, for example, which would do for that phenomenon what chemistry has done for combustion—an explanation which would tell us in physical terms, and without residue, what the experience of colour perception is. Philosophical analyses of the distinguishing features of the mental that are designed to get us over this hurdle generally involve implausible forms of reductionism, behaviouristic in inspiration. The question is whether there is another way of bringing mental phenomena into a unified conception of objective reality without relying on a narrow standard of objectivity which excludes everything that makes them interesting, (p. 1)
Searle believes he has a good physicalist answer—indeed, in its general form the only possible answer—to Nagel’s question. It goes like this: One can accept the reality of conscious mental life without lapsing into dualism, which (he thinks) would be tantamount to abandoning 400 years of cumulative scientific achievement. Consciousness is a biological process as physical as digestion, caused by low-level neurophysiological events in the brain such as neuron firings and the like. But consciousness is not a separate something, over and above the brain action; the causality is not of the boy-hits-ball-then-ball-breaks-window sort. Rather, consciousness emerges, in roughly the same way that solidity emerges when water freezes upon reaching a certain critical temperature. It is a system-level property of the brain. And this is not epiphenomenalism, the idea that consciousness is real but ineffectual; like the solidity of ice or other emergent physical properties, consciousness has causal consequences, causal powers.
Although consciousness is thus at least in principle causally reducible, it is not ontologically reducible; for this would involve showing that what appears to us as consciousness is in reality something else, but in the case of consciousness itself, unlike the apparent rising and setting of the sun, the appearance is the reality. For Searle the CTM was on the right track, insofar as it was a materialist theory of the mind, but it was not nearly materialistic enough. To imagine that the mind can be instantiated in computer programs of any kind is to entertain a kind of residual dualism. The brain generates mind and consciousness by physical, causal mechanisms inherent in our biology, and we need to respect that unique biological reality. This does not entail that mind and consciousness could not appear in other places, for example in some sort of suitably constructed artifact; but in order for this to occur, the physical substrate of the artifact would need to embody causal powers at least equivalent to those inherent in the brain itself.
That summarizes the central theoretical part of Searle’s position. He also goes on to characterize the phenomenology of consciousness in ways that are strikingly reminiscent of James (1890b), and considers in general terms how best to pursue its scientific study. Clearly, given his theoretical commitments, the main order of business is to study the associated neurophysiology. Searle (1997) singled out for special praise neuroscientists such as Gerald Edelman and Francis Crick who were in effect carrying out this program. His admiration for real neuroscience is certainly justified, and it is apparently reciprocated; in a recent paper in the Annual Review of Neuroscience he again summarizes his theoretical views regarding consciousness and offers a variety of additional suggestions, both shrewd and detailed, for further neuroscientific investigation (Searle, 2000).
To summarize: In my view John Searle deserves enormous credit for having almost single-handedly altered the terms of the mind-brain controversy. On the negative side, his destructive critique of the CTM should eliminate from contention a family of views that have dominated the past half-century, especially in psychology. On the positive side, his effective advocacy of “biological naturalism” has focused attention clearly and sharply on the position that I believe represents the ultimate development, the necessary culmination, of a conventional physicalist approach to the mind. The fundamental question before us now reduces starkly to this: Can everything we know about the mind be explained in terms of brain processes?
First, briefly, some conceptual issues. I am not at all persuaded that Searle has succeeded in making meaningful distinctions between his philosophical views and those of a host of others who would describe themselves as advocates of token identity or even property dualism. In particular, his abstract account of the “emergence” of consciousness strikes me as verging upon intellectual sleight of hand. Can mental properties simply be stipulated into physicality in this way? The analogies seem faulty. In the water/ice situation, for example, both ends of the causal relation are indisputably physical things that we know how to observe and measure in conventional physical ways. But that is exactly what is at issue in the relation of brain to mind, as Nagel pointed out.
Related to this, and also captured by Nagel’s remarks, we do not in fact have anything even remotely resembling a full causal account of consciousness, let alone an account that we can understand in the way we understand the freezing of water. Intelligibility of causal accounts is surely something we would like to have, but in this case it seems singularly difficult to imagine how we could possibly get one. I will leave aside here, however, the more philosophic issue whether we should require that a satisfactory causal account be intelligible, and focus instead on the prior empirical question whether we can get one at all.
For Searle it is virtually axiomatic, a given, that brain processes causally generate every detail of our conscious mental life. Throughout his writings he characterizes this as a demonstrated fact, something we know with complete certainty and beyond any possibility of doubt. In the discussion of his paper in Bock and Marsh (1993), for example, he candidly exclaims “frankly, I just can’t take it seriously that my point of view, that brain processes cause consciousness, is denied!” (p. 77).
The vast majority of contemporary neuroscientists and psychologists undoubtedly would agree with Searle in accepting without hesitation this basic premise, although many would perhaps question, as I do, details of his account of the emergence, and his confidence that we will be able to achieve a biological theory that is both adequately complete and theoretically satisfying.20 As noted in the Introduction, the assumption that brain generates mind—that the mind is what the brain does, full stop—seems plausible in light of much prior scientific experience, and has generally served us well as a working hypothesis. It will undoubtedly continue for most scientists to do so. And this is not a bad thing; for as Dupuy (2000) remarked: “The only way to prove the falsity of materialism is to give it every benefit of doubt, to allow it to push forward as far as it can, while remaining alert to its missteps, to the obstacles that it encounters, and, ultimately, to the limits it runs up against. Any other method—fixing its boundaries in advance, for example—is bound to fail” (p. 25).
I agree strongly with this view. I also appreciate that the human brain is a fantastically complex biological system which undoubtedly harbors layer upon layer of neurophysiological mechanism still to be unraveled by our steadily deepening scientific inquiry. New discoveries are constantly being made, some of which profoundly enrich our appreciation of the brain’s resources. For example, the old idea that neurons communicate exclusively by means of direct and specific synaptic connections has been enriched by discovery of complementary “volume-conduction” effects caused by diffusion of neurotransmitters through the local extracellular space (Agnati, Bjelke, & Fuxe, 1992). Again, neuroscientists had long believed that the dendrites which make up the input arbor of a typical cortical neuron are essentially passive cables; but in recent years we have discovered that they are enormously more complex and contribute actively to the overall electrical behavior of the cell (see, e.g., the special issue of Science, October 27, 2000). Similarly, it had long been believed that glial cells provide only structural and metabolic support for the really important elements, the neurons (which they greatly outnumber), but it has recently been demonstrated that by scavenging potassium ions from the extracellular space, glia can also directly modify the electrical behavior of local neural populations (Kohn, Metz, Quibrera, Tommerdahl, & Whitsel, 2000; Kohn, Metz, Tommerdahl, & Whitsel, 2002).21
Other examples could readily be provided, but the point should already be clear: No one can prophesy with certainty how far or even in precisely what directions our evolving understanding of brain mechanisms may lead. Nevertheless, and unlike Dupuy (2000), I believe that sufficient information is already in hand to demonstrate that biological naturalism as currently conceived is not only incomplete but false as a theory of the mind.
At this point we make the empirical turn that is the central and distinctive contribution of this book. I will begin by outlining the conceptual framework that underlies our presentation. Imagine if you will two complex streams of activity flowing through time in parallel, one consisting of your conscious mental experience, the other of the myriad physiological processes going on in your brain. Imagine further, even though it is scarcely feasible in practice, that we could divide both streams individually into a sequence of states, and in such a way that the mental states correspond to the brain states. Suppose still further, and again counterfactually, that perfect 1:1 correspondence between the two sequences has been established empirically. We have discovered the Holy Grail of neuroscience and psychology, the neural correlates of consciousness.
All the traditional philosophical positions on the mind-body problem arise from different ways of interpreting this correlation. The current mainstream consensus is that brain processes generate or constitute mental processes, but this is not the only possible interpretation.
Note first that even perfect correlation would not necessarily entail identity. For example, like all other mammals we have both hearts and lungs, but hearts and lungs are not identical. It remains at least conceptually possible that minds and brains are distinct, though functionally closely linked.
The fundamental scientific strategy for sorting out the causal structure underlying observed correlations is of course to attempt to determine what influences what by manipulating the two sides of the correlation. It is evident, and everyone agrees, that physical events such as sitting on a tack can influence mental events. But what about the other way around? It seems equally obvious, naively, that mental events cause physical events, too. For example, as Searle likes to say, I decide to raise my arm and it goes up. But there is a hitch here, an asymmetry in the causal structure. Physicalists can normally reply, and do in their preferred fashion, that the causality in such cases resides not in the mental per se but in its physical equivalents or accompaniments. In sum, we can cleanly, simply, and directly manipulate the physical side of the correlation, but not so the mental, at least under ordinary conditions.
Nevertheless, we will argue, the correlations we actually observe are not simply incomplete but demonstrably imperfect. There exist certain kinds of empirically verifiable mental properties, states, and effects that appear to outstrip in principle the explanatory potential of physical processes occurring in brains.22 Facts of this sort, moreover, can often be accommodated more naturally within an alternative interpretation of the mind-brain correlation, one already developed in abstract form by William James (1898/1900).
James pointed out that to describe the mind as a function of the brain does not fully specify the character of the functional dependence. Physiologists routinely presume that the role of the brain is productive, the brain generating the mind in something like the way the tea kettle generates steam, or the electric current flowing in a lamp generates light. But other forms of functional dependence exist which merit closer consideration. The true function of the brain might for example be permissive, like the trigger of a crossbow, or more importantly, transmissive, like an optical lens or a prism, or like the keys of a pipe organ (or perhaps, in more contemporary terms, like the receivers in our radios and televisions).
More generally, one can at least dimly imagine some sort of mental reality, which in James’s view might be anything from a finite mind or personality to a World Soul, that is closely coupled to the brain functionally but somehow distinct from it. Within this basic framework James himself spoke variously of the brain as straining, sifting, canalizing, limiting, and individualizing that larger mental reality existing behind the scenes. He also quoted approvingly Schiller’s (1891/1894, pp. 293, 295) characterization of matter as “an admirably calculated machinery for regulating, limiting, and restraining the consciousness which it encases…. Matter is not that which produces Consciousness, but that which limits it, and confines its intensity within certain limits” (James, 1898/1900, pp. 66–67). James also explicitly portrayed the brain as exerting these various effects in a manner dependent on its own functional status, and linked this idea to Fechner’s conception of a fluctuating psychophysical threshold (pp. 24, 59–66).
Much can immediately be said in favor of such a picture, James then argued. It is in principle compatible with all of the facts conventionally interpreted under the production model, and however metaphorical and incomprehensible it might at first seem, it is in reality no more so than its materialist rival. It also has certain positive superiorities. In particular, it appears potentially capable of explaining various additional facts such as those being unearthed by F. W. H. Myers and his colleagues in psychical research (pp. 24–27).
In sum, “transmission” or “filter” models are logically viable, and they should rise or fall in the usual scientific way in terms of their capacity to accommodate the available empirical evidence.23
The remainder of this book consists primarily of a systematic though necessarily incomplete marshaling of evidence and argument supporting filter models in general as an abstract class. Only in Chapter 9, after all this material has been put on the table, will we attempt to identify in greater detail the specific forms we think such theories might take. To complete the present chapter, I will next set forth in brief compass a catalog of types of mental phenomena that appear especially resistant to explanation in conventional mechanistic, biological terms. Most though not all of these phenomena will be discussed in detail in the chapters that follow.
Like James (1898/1900) and McDougall (1911/1961), among many others, I will immediately appropriate the entire body of evidence for psi phenomena in service of our central thesis. Such phenomena by definition involve correlations occurring across physical barriers that should be sufficient, on presently accepted principles, to prevent their formation (Broad, 1962). This occurs for example when a subject successfully identifies randomly selected targets concealed in opaque envelopes, or displayed in a remote location. It is not difficult to set up controlled experiments of this sort and to evaluate their outcomes using rigorous statistical procedures. As stated in the Introduction, and documented in our Appendix, a considerable amount of work has been carried out along these lines, with results more than sufficient to convince me and the other authors of this book that the sheer existence of the basic input/output phenomena—ESP and PK, or in the more theory-neutral terminology of Thouless and Wiesner (1947), “psi”—is an inescapable scientific reality with which we must somehow come to terms.
In this light the anti-psi polemic recently advanced by psychologist/philosopher and long-time skeptic Nick Humphrey (1996) is particularly startling. Throughout his book Humphrey alludes to a supposed killer argument that he will later deploy to demonstrate the impossibility of psi. When we finally get there (Chapter 26), the argument turns out to be that he cannot imagine any possible scenario under which ostensible psi effects could be achieved by some combination of known physical mechanisms. Therefore the reported effects cannot and do not happen, Q. E. D. But whether we like it or not, such effects do happen, as a matter of empirical fact (see the Introduction and the Appendix). That is the whole point, and what makes the phenomena theoretically interesting in the first place! Humphrey’s “argument” amounts in my opinion to little more than an expression of his deeply felt wish that the phenomena should simply go away. In this he is of course adopting a strategy that has been widely practiced by contemporary scientists and philosophers.24
Psi phenomena in general are important because they provide examples of human behavioral capacities that seem extremely difficult or impossible to account for in terms of presently recognized computational, biological, or physical principles. Even more important for our purposes, however, is a further body of evidence suggestive of post-mortem survival, the persistence of elements of mind and personality following bodily death. It is simply false to declare, as does Paul Churchland (1988), that “we possess no such evidence” (p. 10). We in fact possess a lot of such evidence, much of it of high quality (see the Appendix). Ironically, the primary threat to the survivalist interpretation of this evidence arises not from considerations of evidential quality, but from the difficulty of excluding alternative explanations based upon paranormal interactions involving only living persons.
Quite apart from any personal or theological interests readers may bring to this subject, it should be evident that post-mortem survival in any form, if it occurs, demonstrates the presence of a fundamental bifurcation in nature, and hence the falsehood of biological naturalism. We will touch upon various facets of the survival evidence later in the book and summarize our collective sense of the empirical status of the problem in Chapter 9.
Evidence for the occurrence of psi phenomena in general and post-mortem survival in particular thus plays an important though largely implicit role in our overall argument. Our efforts in this book will be amply rewarded if they lead scientifically-minded readers to take these subjects more seriously than they otherwise might. There are many other kinds of evidence, however, that point in the same general direction.
Under this heading comes a variety of phenomena especially suggestive of direct mental agency in the production of physiological or even physical effects. Chapter 3 discusses many such phenomena in detail, but I will give a few examples here to capture their flavor.
Placebo effects and related kinds of psychosomatic phenomena have long been informally recognized and are now widely accepted, but they were accepted by modern medical science only grudgingly, as new mechanisms of brain-body interaction came to light that seemed potentially capable of accounting for them. There remain many types of kindred phenomena, however, that pose progressively more severe challenges to explanation in such terms.
Myers, for example, was interested in hysterical “glove anesthesias,” in which a patient loses sensation from the skin of a hand in the absence of organic lesion. In such cases the anesthetic skin region typically corresponds only to a psychological entity, the patient’s idea, in complete disregard of the underlying anatomical organization. At the same time, curiously, something in the patient remains aware of the afflicted region and protects it from injury.
Related phenomena have been reported in the context of hypnosis. For example, highly suggestible persons who can vividly imagine undergoing an injurious circumstance such as receiving a burn to the skin sometimes suffer effects closely analogous to those that the physical injury itself would produce, such as a blister. More rarely, the correspondence between the hypnotic blister and its imagined source extends even to details of geometric shape, details which appear especially hard to account for in terms of known mechanisms of brain-body interaction. A closely related and well-documented phenomenon is that of “stigmata,” in which fervently devout or pious believers in Christ develop wounds analogous to those inflicted during his crucifixion. The injuries to the skin are again localized and specific in form, and they differ in locus and character in accordance with their subjects’ varying conceptions of Christ’s own injuries. Similarly dramatic phenomena have occasionally been documented in psychiatric patients in connection with their recall of prior physical trauma.
The conventional expectation, of course, is that even the most extreme of the phenomena just mentioned can ultimately be explained in terms of brain processes. Continuing allegiance to this expectation, despite the explanatory difficulties, is undoubtedly encouraged by the fact that the phenomena described so far all involve effects of a person’s mental states on that person’s own body. Even more drastic explanatory challenges are posed, however, by additional and related phenomena in which one person’s mental state seems to have directly influenced another person’s body. Such phenomena include “maternal impressions” (birthmarks or birth defects on a newborn that correspond to an unusual and intense experience of the mother during the pregnancy), distant healing, experimental studies of distant mental influence on living systems, and cases in which a child who claims to have memories of the life of a deceased person also displays unusual birthmarks or birth defects corresponding closely with marks (usually fatal wounds) on the body of that person. In addition, there has been a considerable influx since Myers’s time of other experimental evidence demonstrating the reality of psychokinesis (PK), which by definition involves direct mental influence on the physical environment.
Chapter 3 presents selective and focused discussions of phenomena of these various types, emphasizing their strong association with factors such as emotion, strong beliefs, and unusually vivid mental imagery, and drawing out their implications for an adequate theoretical picture of consciousness, volition, and the mind-body problem. Within-subject phenomena such as placebo effects, hypnotically induced blisters, and stigmata are also more carefully situated in relation to modern developments in “psychoneuroim-munology” and mind-body medicine. The common threads of the chapter are, first, to point out the many, varied, and well-documented phenomena of extreme psychophysical influence for which no conventional physicalist explanation is presently available, or in some cases even seems possible; and, second, to point out the theoretical continuity between normal, conscious volitional acts and these less common phenomena that suggest unconscious (or subliminal) volitional activity, sometimes of wider scope than conscious volition itself. The chapter also provides striking examples of the sometimes pathological interplay between scientific theories and scientific “fact.”
A number of well-documented psychological phenomena involve levels of detail, precision, or logical depth that seem hard to reconcile with what can be achieved by a fundamentally analog mechanism operating in a statistical way with neural components of low intrinsic precision and reliability. I will give three examples of the sort of thing I have in mind.
The first comes from a case of “automatic writing” observed by James (1889). The subject wrote with his extended right arm on large sheets of paper, his face meanwhile buried in the crook of his left elbow. For him to see what he was doing was “a physical impossibility.” Nevertheless, James continues: “Two or three times in my presence on one evening, after covering a sheet with writing (the pencil never being raised, so that the words ran into each other), he returned to the top of the sheet and proceeded downwards, dotting each i and crossing each t with absolute precision and great rapidity” (p. 44).
This extraordinary performance illustrates two features that have often appeared together in the substantial but neglected scientific literature dealing with automatic writing (Koutstaal, 1992; Stevenson, 1978): The subject is in an altered state of consciousness, and the motor performance, itself extraordinary, is apparently guided by an extremely detailed memory record, an essentially photographic representation of the uncompleted page.
The latter property relates to the phenomenon of eidetic imagery, my second example, study of which was revived in modern times by the Habers (see Obler & Fein, 1988, for an overview). Probably the most dramatic demonstration of its psychological reality has been provided using Julesz random-dot stereograms (Stromeyer, 1970; Stromeyer & Psotka, 1970; see also Julesz, 1971, pp. 239–246). These are essentially pairs of computer-generated pictures, each of which by itself looks like a matrix of randomly placed dots, but constructed in such a way that when viewed simultaneously (by presentation to the two eyes separately) a visual form emerges in depth. Stromeyer and Psotka adapted this technology to their aims by presenting pictures of this type to the eyes of their single subject, a gifted female eidetiker, at different times, ultimately as much as three days apart. Under these conditions, the subject could only extract the hidden form if she could fuse current input to one eye with an extremely detailed memory-image of previous input to the other eye. Remarkably, she was able to succeed under a wide variety of increasingly severe demands. The original stereograms, for example, were 100 x 100 arrays, but she ultimately succeeded under double-blind conditions with arrays as large as 1000 x 1000, or a million “bits,” viewed up to four hours apart.
These results were understandably shocking to many psychologists, who sought to escape their force by pointing to the dependence on a single gifted subject and the absence of replications (R. Herrnstein, personal communication, October, 1972). At least one successful replication has subsequently occurred, however. Specifically, Crawford, Wallace, Nemura, and Slater (1986) demonstrated that their highly hypnotizable subjects were able to succeed with the small (100 x 100) stereograms, but only when they were hypnotized. Moreover, the literature already contains many additional examples of prodigious memory. Stromeyer and Psotka themselves mention the mnemonist intensively studied by Luria (1968) and the case of the “Shass Pollaks,” who memorized all 12 volumes of the Babylonian Talmud (Strat-ton, 1917). Sacks (1987, chap. 22) has reported a similar case of a person who among other things knew by heart all nine volumes and 6,000 pages of Grove’s Dictionary of Music and Musicians. Other examples could easily be cited. Prodigious memory of this sort appears to be a real psychological phenomenon.
Third in this group is the whole family of “calculating prodigies.” Of special interest is the “savant syndrome,” often associated with infantile autism, in which islands of spectacular ability appear in the midst of generalized mental disability (Obler & Fein, 1988; Treffert, 1989). The abilities are of many types, but almost invariably involve prodigious memory. The depth of the problems they pose for brain theory is exemplified by the case of “The Twins,” also described by Sacks (1987). These individuals, unable to perform even simple additions and subtractions with any accuracy, nonetheless proved able to generate and test prime numbers “in their heads.” Sacks was able to verify the primacy up to 10 digits, but only by means of published tables, while the twins themselves went on happily exchanging ostensibly prime numbers of even greater length, eventually reaching 20 digits. Sacks makes the intriguing suggestion that they seem not to be literally calculating these enormous numbers, but discovering them by navigating through some vast inner iconic “landscape” in which the relevant numerical relations are somehow represented pictorially. The twins themselves of course cannot say how they do it.
Phenomena of the sorts described in this section look hard to explain in terms of brain processes. The most serious attempt to do so known to me (Snyder & Mitchell, 1999) is in fact devoid of specific neural mechanisms. Its central argument is rather that early-stage brain processes like those subserving visual perception, for example, must also be rather savant-like in terms of their speed, precision, and informational capacity; what is unusual about savants, therefore, may consist merely in their access to these mechanisms. This “explanation” of course presupposes a positive answer to the fundamental question at issue, whether the brain alone can accomplish any of these things, including perceptual synthesis itself (see below).
I will make just one further observation before leaving this fascinating and challenging subject. The biocomputational approach leads to one further expectation that could readily be tested using brain imaging methods but to my knowledge has not. As proved by von Neumann (1956), the only practical way to get increased arithmetical depth and precision out of individually unreliable computing elements is to use more of them. Although I do not know how to quantify this in a rigorous way, the biocomputational perspective clearly implies that calculating prodigies must use very large portions of their brains in very abnormal ways to achieve the observed effects. The cognitive losses that often accompany savant skills could perhaps be a reflection of such substitution, but we must remember that savant-type skills sometimes also occur in geniuses such as the mathematicians Gauss and Ampére (see also Obler & Fein, 1988, chap. 23).
The previous section focused on phenomena such as prodigious memory that appear potentially incompatible with the physical properties of the brain considered as a kind of computing device. Problems also arise, however, in regard to memory in its more familiar and everyday forms. Here I provide a brief sketch of the relevant issues, which are discussed in depth in Chapter 4.
Memory is increasingly recognized as central to all human cognitive and perceptual functions, yet we remain largely ignorant of where and in what forms our past experience is stored and by what means it is brought to bear upon the present. Generations of psychologists and neurobiolo-gists have taken it as axiomatic that all memories must exist in the form of “traces,” physical changes produced in the brain by experience, but there has been little progress toward scientific consensus on the details of these mechanisms despite many decades of intensive research.
Significant progress has recently been made, to be sure, in regard to “learning” and “memory” in simple creatures such as the sea-slug (Aply-sia), and more generally in regard to what might be called “habit memory” (Bergson, 1908/1991), the automatic adjustments of organisms to their environments. But these discoveries fall far short of providing satisfactory explanations of the most central and important characteristics of the human memory system, including in particular our supplies of general knowledge (semantic memory) and our ability to recall voluntarily and explicitly our own past experience (autobiographical or episodic memory). Furthermore, recent functional neuroimaging studies, although generating vast amounts of “data,” have yielded little if any progress toward a comprehensive and coherent account of memory based on trace theory.
Meanwhile, deep logical and conceptual problems have been identified in the root notion of the memory “trace” itself, as well as in allied components of current mainstream doctrine such as “information” and “representation.” Most challenging of all to mainstream views is the substantial body of evidence that has accumulated—much of it since Myers’s original efforts along these lines—suggesting that autobiographical, semantic, and procedural (skill) memories sometimes survive bodily death. If this is the case, memory in living persons presumably exists at least in part outside the brain and body as conventionally understood. At the very least this evidence raises profound and interesting issues of both philosophical and practical importance in regard to our criteria and procedures for determining personal identity.
I believe that the evidence and issues discussed in Chapter 4 collectively suggest the need for a radical reconceptualization of human memory along lines suggested a century ago by Myers, James, and Bergson. We will return to this theme in Chapter 9.
Phenomena catalogued under this heading involve what looks like multiple concurrent engagement, in potentially incompatible ways, of major cognitive skills (linguistic skills, for example) and the corresponding brain systems. Let me explain.
The current mainstream view pictures the mind or “cognitive system” as a hierarchically ordered network of subprocessors or “modules,” each specialized for some particular task and corresponding (it is hoped) to some particular brain region or regions. Leaving aside major issues regarding the details of its specification, this picture seems broadly consistent with the overall manner in which our minds seem normally to operate. Our basic way of doing things, that is, is essentially one at a time in serial fashion. Although psychologists recognize that with suitable training people can do more things simultaneously than they customarily suppose, this generalization applies mainly to relatively divergent things, and conspicuously fails as the simultaneous tasks become more complex and similar (Baars, 1997; Neisser, 1976; Pashler, 1998).
Nevertheless, a large body of credible evidence, some dating back to the late 19th century, demonstrates that additional “cognitive systems,” psychological entities indistinguishable from full-fledged conscious minds or personalities as we normally understand these terms, can sometimes occupy the same organism simultaneously, carrying on their varied existences as it were in parallel, and largely outside the awareness of the primary, everyday consciousness. In essence, the structure that cognitive psychology conventionally pictures as unitary, as instantiated within and identified with a particular organization of brain systems, can be functionally divided—divided, moreover, not horizontally, leading to isolation of the normal cognitive capacities from each other, but vertically, leading to the appearance of what seem to be two or more complete cognitive systems each of which includes all of the relevant capacities. Even worse, it sometimes happens that one of these “multiple” or “alter” personalities appears to have direct access to the conscious mental activity of one or more others, but not vice versa.
Investigation of such phenomena, and more generally of “consciousness beyond the margin” (E. Taylor, 1996) was central to the early work of Myers and James among many others. Two brief examples may serve to convey a more concrete initial sense of the character of these phenomena.
The first comes from Myers (HP, vol. 2, pp. 418–422), who quoted from a report by Oxford philosopher F. C. S. Schiller on automatic writing produced by his brother. As is characteristic of this genre of automatisms, the writer typically had no consciousness of the content of his writing, which went on continuously while he was fully and consciously engaged in some other activity such as reading a book or telling a story. Of particular relevance here, however, were occasions on which he wrote simultaneously with both hands and on completely different subjects, one or the other of these streams of writing also sometimes being in mirror-image form.
The second example is the case of Anna Winsor, described both by James (1889) in his report on automatic writing and by Myers (HP, vol. 1, pp. 354–360). The case was protracted and bizarre, but only superficially resembles the neurological “alien hand” (Dr. Strangelove) syndrome. Its central feature is that the patient, Anna, at a certain point lost voluntary control of her right arm, which was taken over by a distinctive secondary personality. This personality, whom Anna herself named “Old Stump,” was benign, often protecting Anna from her pronounced tendencies toward self-injury. As in the case of Schiller’s brother, Stump also typically wrote or drew while Anna herself was occupied with other matters. But Stump also continued writing and drawing even when Anna was asleep, and sometimes in total darkness. This secondary personality also remained calm and rational during periods when Anna herself was feverish and delusional, and it manifested knowledge and skills which Anna did not possess.
The enormous literature on these subjects is reviewed systematically, and its implications discussed, in Chapter 5. The chapter specifically argues for the psychological reality of the phenomenon of co-consciousness (as distinguished from unconscious cerebration and alternating consciousness), which is fundamental both to Myers’s own theoretical framework and (as discussed in Chapter 8) to James’s later application of that framework to problems in religion and philosophy.
Under this heading I will briefly address two interrelated problems. The first and narrower is the so-called “binding” problem. This problem emerged as a consequence of the success of contemporary neuroscientists in analyzing sensory mechanisms, particularly in the visual system. It turns out that different properties of a visual object such as its form, color, and motion in depth are handled individually by largely separate regions or mechanisms within the visual system. But once the stimulus has been thus dismembered, so to speak, how does it get back together again as a unit of visual experience?
Only one thing is certain: The unification of experience is not achieved anatomically. There are no privileged places or structures in the brain where everything comes together, either for the visual system by itself or for the sensory systems altogether. McDougall (1911/1961) was already fully aware of this, and used it as a cornerstone of his argument against materialist accounts of the mind. In his view, the evident disparity between the multiplicity of physiological processes in the brain and the felt unity of conscious experience could only be resolved in physicalist terms by anatomical convergence, and since there is no such convergence, physicalism must be false.25
McDougall’s original argument, although ingenious, relied upon the faulty premise that the only possible physical means of unification must be anatomical. All current neurophysiological proposals for solving the binding problem are instead functional in nature. The essential concept common to all of them is that oscillatory electrical activity in widely distributed neural populations can be rapidly and reversibly synchronized in the “gamma” band of frequencies (roughly 30–70 Hz), thereby providing a possible mechanistic solution to the binding problem (von der Malsburg, 1995).26
A great deal of sophisticated experimental and theoretical work over the past 20 years has demonstrated that such mechanisms do in fact exist in the nervous system, and that they are active in conjunction with normal perceptual synthesis.27 Indeed, Searle’s doctrine of biological naturalism has now crystallized neurophysiologically in the form of a family of “global workspace” theories, all of which make the central claim that conscious experience occurs specifically and only in conjunction with large-scale patterns of gamma-band oscillatory activity linking widely separated areas of the brain (Crick, 1994; Dehaene & Naccache, 2001; Edelman & Tononi, 2000; A. K. Engel, Fries, & Singer, 2001; W. J. Freeman, 2000; Llinás, Rib-ary, Contreras, & Pedroarena, 1998; Newman & Baars, 1993; Singer, 1998; Varela, Lachaux, Rodriguez, & Martinerie, 2001).28
The neurophysiological global workspace, however, cannot be the whole story. A sizeable body of recent evidence demonstrates that organized, elaborate, and vivid conscious experience sometimes occurs under physiological conditions, such as deep general anesthesia and cardiac arrest, which preclude workspace operation. Experiences of this sort fall under the more general heading of “near-death experiences” or NDEs, which are discussed in detail in Chapter 6 together with related phenomena such as “out-of-body experiences” (OBEs) and lucid dreams. In short, it appears to us that McDougall was right after all, albeit for the wrong reason. In effect, we will argue, recent progress in mainstream physicalist brain theory has provided new means for its own falsification as a complete account of mind-brain relations.
Availability of this emerging evidence emboldens me to make some further and even more contentious remarks regarding the larger problem of perceptual synthesis, and the direction in which things seem to me to be moving.
It is an historical fact that mainstream cognitive psychology has always tended on the whole to try to solve its problems “on the cheap,” with as little reference as possible to what all of us experience every day as central features of our conscious mental life. The early workers in “mechanical translation,” for example, imagined that they could do a decent job simply by constructing a large dictionary that would enable substitution of words in one language for words in the other. This approach failed miserably, and we were slowly driven, failed step by failed step, to the recognition that truly adequate translation presupposes understanding, or in short a full appreciation of the capacities underlying the human use of language.
A similar evolution is underway in regard to perceptual theory. Most of the work to date has taken a strongly “bottom-up” approach, along lines formulated in the seminal book of Marr (1982). This school views perceptual synthesis as a kind of exhaustive calculation from the totality of input currently present at our sensory surfaces. Machine vision and robotics, for example, necessarily took this approach, and even in neuroscience it seemed to make sense to start with the most accessible parts of the perceptual systems—the end organs and their peripheral connections—and work our way inward. The great sensory systems themselves—vision, audition, somato-sensation, and so on—were also presumed to operate more or less independently, and were in fact typically studied in isolation.
A separate tradition dating back at least to Kant and the early Gestalt theorists, and carried forward into the modern era by psychologists such as Neisser (1967, 1976), has been sensitive to the presence of “top-down” influences, both within and between sensory modalities. Although a few perceptual subsystems (such as those that engender the Müller-Lyer and other incorrigible visual illusions) may be truly autonomous or “cognitively impenetrable” in the sense of J. Fodor (1983), these seem to be isolated and special cases. A very different overall picture of perceptual synthesis is currently emerging in which top-down influences predominate. On this view perceptual synthesis is achieved not from the input, but with its aid. This is necessarily the case for example in regard to ambiguous figures, where the stimulus information itself is not sufficient to determine a uniquely “correct” interpretation. More generally, we routinely ignore information that is present in the input and supply information that is not, speed-reading providing a characteristic example.29 Something within us, a sort of cosmo-genic, world-generating, or virtual-reality system, is continuously updating and projecting an overall model of the perceptual environment and our position within it, guided by very limited samplings of the available sensory information (Simons & Chabris, 1999; Tart, 1993).
As in the case of understanding spoken or written language, an enormous amount of general knowledge is constantly mobilized in service of this projective activity, which freely utilizes any and all relevant sensory data available to it. Top-down and cross-modal sensory interactions have recently been recognized as the rule rather than the exception in perception (A. K. Engel et al., 2001; Shimojo & Shams, 2001). Neuroscientist Rodolfo Llinás and his co-workers have even advanced the view, which I believe is profoundly correct, that dreaming, far from being an odd and incidental part of our mental life, represents the fundamental form of this projective activity. Ordinary perceptual synthesis, on this inverted view of things, amounts to oneiric (dreamlike) activity constrained by sensory input (Llinás & Paré, 1996; Llinás & Ribary, 1994). Hartmann (1975) had proposed similar ideas in regard to hallucinatory activity more generally, with dreaming included. On his view such activity is again a ubiquitous and fundamental feature of our mental life, and the critical question is not “why do we sometimes hallucinate?” but rather “what keeps us from hallucinating most of the time?” The answer, he thought, lies in inhibitory influences somehow exerted by the brain activity that accompanies ongoing perceptual and cognitive functions of the ordinary waking sorts. Similar arguments for the primacy and importance of this cosmogenic capacity have more recently been advanced by Brann (1991) and Globus (1987).30
So far so good, but where exactly is the “top,” the ultimate source of this projective activity? The mainstream neuroscientists who have already recognized its existence invariably presume that it arises entirely within the brain itself. Evidence such as that assembled in Chapter 6, however, like the more direct evidence of post-mortem survival, strongly suggests that it originates outside the brain as conventionally understood. We will return to this in Chapter 9.
In the Introduction we quoted Edgar Wind’s guiding principle, that “the commonplace may be understood as a reduction of the exceptional, but the exceptional cannot be understood as an amplification of the commonplace.” That principle applies with particular force and poignancy, I think, to the topic of this section. Any scientific theory of personality and cognition truly worthy of the name surely must help us to understand this humanly vital topic, but by this standard we have so far made distressingly little progress. The reason, in my opinion, is that for the most part we have violated Wind’s principle by trying to understand the exceptional—real genius, in its fullest expressions—as an amplification of the commonplace—”creativity,” as we find it in random samples of undergraduates and the like.
Myers and James consciously and deliberately approached the subject from the other end, and in connection with the enlarged conception of human personality which they were struggling to articulate. In Chapter 7 we discuss genius from this point of view, describing Myers’s account in some detail and situating it in relation to the main trends in contemporary creativity research. Focusing primarily on the creative process and creative personality structure, we argue that Myers anticipated most of what has been good in more recent work, while also accommodating in a natural way a variety of additional phenomena—including psychological automatisms and secondary centers of consciousness, altered states of consciousness, unusual forms of symbolic thinking, and psi—that are inescapably bound up with this topic but scarcely touched upon in contemporary mainstream accounts. We also show that various expectations flowing from Myers’s account of genius have been strongly confirmed by more recent empirical observations.
Experiences of this type lie at the core of the world’s major religious traditions and have continued to occur throughout history and across cultures. Their existence as a distinctive and important class of psychological phenomena can scarcely be denied. Yet they have largely been ignored by mainstream psychology and neuroscience, and generations of reductionist clinical psychologists and psychiatrists have consistently sought to devalue and pathologize them. Even when acknowledging that such experiences are often life-transforming and self-validating for those who have them, the historically standard epistemological approaches in psychology and philosophy—beginning with William James in his Varieties of Religious Experience—treat them as purely subjective events having authority only for those who experience them, and thus deny their objective significance and the testability of the associated truth-claims. However, a large though scattered literature testifies to the occurrence in such experiences, or in individuals who have them, of genius-level creativity and many other unusual empirical phenomena of the sorts discussed in this book. Mystical-type states of consciousness are also now known to be at least partially reproducible by pharmacological means (psychedelics), and they can also be induced by protracted self-discipline involving transformative practices such as the various forms of meditation. A more objective, informed, and sympathetic appraisal of mystical experience finds within it much additional support for a Myers-type theory of human personality, and many new opportunities for empirical research. Chapter 8 will develop these themes in detail.
In this final section I wish to comment briefly on a hornet’s nest of theoretical issues lying at the very core of our mental life, issues that have been the focus of extensive recent debates, especially in the philosophical literature.31 These issues are deep, individually complex, and densely interconnected. What I have to say will necessarily amount to little more than a summary of my own opinions, but I believe that this statement derives support both from earlier sections of this chapter and especially from the chapters that follow.
The crucial point I want to make, especially to my fellow psychologists, is this: Our a priori commitment to a conventional physicalist account of the mind has rendered us systematically incapable of dealing adequately with the mind’s most central properties. We need to rethink that commitment.
Consider first the issue of semantic or intentional content, the “meaning” of words and other forms of representation. Throughout our history, we have tried unsuccessfully to deal with this by “naturalizing” it, reducing it to something else that seems potentially more tractable. An old favorite among psychologists, traceable at least back to Locke and Hume, is that representations work by resembling what they represent, by virtue of some sort of built-in similarity or structural isomorphism. Any hope along these lines was long ago exploded in philosophical circles, as shown for example by Goodman (1972) and by Heil (1981) in his destructive review of J. Fodor (1975). The central move subsequently made by classical cognitive psychology is essentially the semantic counterpart of the prevailing functionalist doctrine in philosophy of mind: Meanings are not to be conceived as intrinsic to words or concepts on this view, but rather as deriving from and defined by the functional role those words or concepts play in the overall linguistic system. Currently, there is great interest in “externalist” causal accounts of this functionalist type. In connectionism, dynamic systems theory, and neuroscience, for example, the “meaning” of a given response, such as the settling of a network into one of its “attractors” or the firing off of a volley of spikes by a neuron in visual cortex, is typically identified with whatever it is in the organism’s environment that produces that response. But this simply cannot be right. How can such an account deal with abstract things, for example, or non-existent things? Responses do not qualify ipso facto as representations, nor signs as symbols. Something essential is being left out. That something, as Searle (1992) so effectively argues, is precisely what matters, the semantic or mental content (see also the detailed discussion of “representations” in Chapter 4).
Closely related to this is the more general and abstract problem of intentionality, the ability of any and all representational forms to be about things, events, and states of affairs in the world. Mainstream psychologists and philosophers have struggled to find ways of making intentionality intrinsic to the representations themselves, but again it just does not and cannot work, because something essential is left out. That something is the user of the representations. Intentionality is inherently a three-way relation involving users, symbols, and things symbolized, and the user cannot be eliminated. As Searle puts it in various places, the intentionality of language is secondary and derives from the intrinsic intentionality of the mind. Searle thus agrees in part with Brentano (1874/1995), for whom intentionality was the primary distinguishing mark of the mental. At the same time, however, Searle ignores the other and more fundamental part of Brentano’s thesis, which is that intentionality cannot be obtained from any kind of physical system, presumably including brains (see Puccetti, 1989, and Dupuy, 2000, for useful discussions of this issue from opposing points of view).
Talk of “users” and the like raises for many contemporary psychologists and philosophers the terrifying specter of the homunculus, a little being within who embodies all the capacities we sought to explain in the first place. Such a result would clearly be disastrous, because that being would evidently need a similar though smaller being within itself, and so on without end. Cognitive modelers seeking to provide a strictly physicalist account of the mind must therefore do so without invoking a homunculus, but they have not succeeded. Often the homuncular aspect is hidden, slipped into a model by its designers or builders and covertly enlisting the semantic and intentional capacities of its users or observers. The semantic theory of J. J. Katz and Fodor (1964), mentioned earlier, provides one example of this, in that the notation they used to designate their “semantic markers” progressively took on characteristics of telegraphic speech as the theory evolved (E. F. Kelly, 1975). Much of the contemporary work on computational modeling of metaphor has similar problems (Chapter 7).
Sometimes, however, the homunculus is more brazenly evident. One example is Marr’s account of vision, which applies computations to the two-dimensional array of retinal input in order to generate a “description” of the three-dimensional world that provided that visual input, but then needs someone to interpret that description (Searle, 1992). Another is Kosslyn’s model of visual imagery, which essentially puts up an image on a sort of internal TV screen, but then needs somebody else to view that image. Draa-isma (2000) and our Chapter 4 identify similar homunculus problems in the context of contemporary memory models.
Particularly in its more blatant forms the homunculus problem has attracted the attention of physicalists such as Dennett (1978), who have sought to remove its philosophic sting. Dennett’s solution is to “discharge” the homunculus by a process of “recursive decomposition.” The basic idea is that the “smart” homunculus appearing at the top of a model can be replaced by progressively larger numbers of less smart homunculi, until we get to a vast bottom layer corresponding to the “hardware” level of computer flip-flops or neuron firings. But as Searle (1992) pointed out, this maneuver fails, because even at the bottom level there has to be something outside the decomposition, a homunculus in effect, that knows what those lowest-level operations mean. Cognitive models cannot function without a homunculus, in short, precisely because they lack what we have—minds, with their capacities for semantics, intentionality, and all the rest built in.
No homunculus problem, however, is posed by the structure of our conscious experience itself. The efforts of Dennett (1991) and others to claim that there is such a problem, and to use that to ridicule any residue of dualism, rely upon the deeply flawed metaphor of the “Cartesian theater,” a place where mental contents get displayed and I pop in separately to view them. Descartes himself, James, and Searle, among others, all have this right; conscious experience comes to us whole and undivided, with the qualitative feels, phenomenological content, unity, and subjective point of view all built-in, intrinsic features. I and my experience cannot be separated in this way.
Finally, I wish simply to record, without argument, my own deepest intuition as to where these issues lead. All of the great unsolved mysteries of the mind—semantics, intentionality, volition, the self, and consciousness—seem to me inextricably interconnected, with consciousness somehow at the root of all.
The consciousness I have in mind is emphatically not that of Chalmers (1996), irreducible but ineffectual, consisting merely of phenomenological properties or “qualia” arbitrarily tacked on to a strong artificial intelligence that does all the cognitive work. Ordinary perception and action are saturated with conceptual understanding, and conceptual understanding is saturated with phenomenological content. Volition too has an intentionality aspect, for as Nietzsche somewhere remarked, one cannot just will, one must will something. Each individual word is in effect “a microcosm of human consciousness” (Vygotsky, 1986/2000, p. 256), and “all meaning is in some way ultimately grounded in being” (Cassirer, 1957, p. 94). And as William James so forcibly argued at the dawn of our science, all of this perceptual, cognitive, and volitional activity somehow emanates from a mysterious and elusive “spiritual self,” which can often be sensed at or behind the innermost subjective pole of our ongoing conscious experience.
Consciousness, in short, far from being a passive epiphenomenon, seems to me to play an essential role—indeed the essential role—in all of our most basic cognitive capacities. I can find no better way of ending this section than simply to stand back and applaud the trenchant conclusion drawn by philosopher E. J. Lowe (1998), which encapsulates my own views and the central contention of this book: “Reductive physicalism, far from being equipped to solve the so-called ‘easy’ problems of consciousness, has in fact nothing very useful to say about any aspect of consciousness” (pp. 121–122).
In regard to the deep theoretical issues broached in the previous section, my views as a psychologist closely parallel those of a small minority of modern philosophical writers including E. J. Lowe, Thomas Nagel, and John Searle. I feel an enormous debt of gratitude to them for so vigorously defending, like James (1890b), the reality and importance of our conscious mental life. I find it altogether astonishing, and predict that it will be found so as well by our intellectual descendants, that so much of our science and philosophy from James to the present has sought—consciously!—to slight or ignore these first-person realities of the mind, and sometimes even to deny that they exist. There is perhaps no better example of the power of theory to blind us to facts.
But that has now all changed. We have come full circle and surely will not turn backward again. The question that now confronts us, as it confronted James, is whether all this richness of our conscious mental life really can be accounted for in terms of physical operations in the brain. Searle himself, of course, is sure that it can, at least in principle. But this is just a pious hope, the latest form of “promissory materialism” (Popper & Eccles, 1977, p. 96). Even philosophers fundamentally sympathetic to Searle’s point of view are not nearly so confident. Consider for example the gloomy conclusions of physicalist Jaegwon Kim (1998):
We find ourselves in a profound dilemma: If we are prepared to embrace reductionism, we can explain mental causation. However, in the process of reducing mentality to physical/biological properties, we may well lose the intrinsic, subjective character of our mentality—arguably, the very thing that makes the mental mental. In what sense, then, have we saved “mental” causation? But if we reject reductionism, we are not able to see how mental causation should be possible. But saving mentality while losing causality doesn’t seem to amount to saving anything worth saving. For what good is the mind if it has no causal power? Either way, we are in danger of losing mentality. That is the dilemma, (p. 237)
In this passage Kim, a determined but scrupulously honest physicalist, has clearly moved within a hair’s breadth of the more skeptical and agnostic Nagel. This surely is a development that ought to worry other physicalists. But whereas their troubles with physicalism arose primarily via conceptual analysis, and have led them to no definite conclusion, ours are primarily empirical in character, and actually falsify biological naturalism. That is the central contention of this book. We believe strongly that in order to get an adequate scientific account of the mind we must be prepared to take seriously all relevant data and to modify as necessary even our most fundamental theoretical ideas. Some of the most relevant kinds of data, however, have been systematically excluded from contemporary scientific and philosophic discussions.
Following chapters, as indicated, will discuss a number of these neglected topics in depth. Let me conclude by briefly recapitulating the argument as it has developed so far: Despite its many significant accomplishments, a century of mainstream scientific psychology has not provided a satisfactory theory of mind, or solved the mind-body problem. Physicalist accounts of the mind appear to be approaching their limits without fully accounting for its properties. The computational theory of the mind has been overthrown, forcing physicalism to retreat into what necessarily constitutes its final frontier, the unique biology of the brain. But this biological naturalism appears destined to fare little better. Some critical properties of mental life can already be recognized as irreconcilable in principle with physical operations of the brain, and others seem likely to prove so as well.
These failures warrant a serious attempt to rethink the entire subject of mind-brain relations from a different perspective—specifically, from the perspective provided by James’s generic “transmission” theory. There is no better way to begin such an effort than by reviewing the extraordinary contributions of F. W. H. Myers, James’s colleague and friend, who provided the most fully developed version so far of a theory of this type.
1. This chapter updates and greatly expands on views more tentatively expressed in E. F. Kelly (1979). I wish to express thanks to Lisette Coly and the Parapsychology Foundation for permission to use parts of that earlier essay. Thanks also to Alan Gauld for especially helpful comments on earlier drafts of this one.
2. The symposium recorded in Bock and Marsh (1993) provides an outstanding example, and one we have tried collectively to emulate.
3. For lucid and accessible discussions of the various philosophic positions briefly canvassed in this section see for example the books of Churchland (1988), Heil (1998), Kim (1998), and Searle (1992).
4. But note that Putnam himself subsequently abandoned functionalism, becoming one of its severest critics; see Putnam (1988/1998).
5. Some observers, however, already resisted this interpretation of the behavioral appearances as truly equivalent to intrinsic purposefulness in the everyday psychological sense; see, for example, Buckley (1968, Part V).
6. Readers interested in following the history of such work can refer to landmark publications such as Feigenbaum and Feldman (1963), Minsky (1968), Schank and Colby (1973), Winograd (1972), and Winston (1975), or for an overview H. Gardner (1985).
7. For further details see E. F. Kelly (1975).
8. Among other things, Weizenbaum was author of the widely known ELIZA program, which simulates a non-directive psychotherapist by means of various simple technical devices that involve no “understanding” whatever. Weizenbaum had been horrified to discover many persons interacting with this program as though it were an actual human being.
9. Fundamentally similar views have also been advanced by writers such as Po-lanyi (1966) and Searle (1992, chap. 8). It is ironic, and perhaps symptomatic, that CTM advocate Robert Harnish (2002, pp. 107–123) reports enthusiastically on SHRDLU without even mentioning Winograd’s defection.
10. But see Brann (1991) and N. J. T. Thomas (1999) for excellent third-party reviews.
11. Some important early landmarks here are books by H. Gardner (1976), Luria (1966), Sacks (1987), and Shallice (1988).
12. In addition to the “macro” technologies described in the text, which operate on scales from the whole brain down to roughly naked-eye-sized parts of it, contemporary neuroscience has developed an impressive arsenal of “micro” technologies, suitable mainly for use in animal studies, that operate down to the cellular or even subcellular level. Other emerging developments at the macro level that look especially promising for future work in humans include transcranial optical imaging and transcranial magnetic stimulation.
13. Most cognitive neuroscientists clearly regard such work as repeatedly and unambiguously confirming the basic mind-brain doctrine of contemporary physical-ism, but many significant technical uncertainties remain regarding procedures for acquisition, processing, and interpretation of imaging data, and some of the available results actually conflict with that doctrine. We will touch upon these issues later in this chapter, as well as in our Chapters 4 and 9.
14. Excellent general introductions are provided by Bechtel and Abrahamsen (2002) and Harnish (2002), and many of the most important early papers are reprinted, with illuminating commentaries, by J. A. Anderson and Rosenfeld (1988).
15. For useful general introductions to dynamic systems see Bechtel and Abra-hamsen (2002, chap. 8), A. Clark (2001), Dupuy (2000), Kelso (1995), Port and van Gelder (1998), and papers by Beer (2000) and van Gelder (1998).
16. Parenthetically, a neglected paper by Heil (1981) argued on very similar grounds that the computationalism of J. Fodor (1975) is fundamentally mistaken.
17. According to Dupuy (2000, pp. 141–142), Kurt Gödel himself specifically stated that irreducibility of semantics to syntax was the “real point” of his incompleteness theorem. See also Penrose (1989, 1994).
18. See Searle (1980) for some choice examples of these reactions. Additional examples of the quasi-religious fervor of hard-core compuphiliacs can be found in Brockman (2003). Michael Grosso informs me that this kind of investment of religiosity in materialism as a philosophic doctrine goes back at least as far as Em-pedocles, who was called “soter” (savior) by his followers.
19. Parenthetically, shifting the effective locus of the computations to a level above (E. L. Schwartz, 1999) or below (Penrose, 1989, 1994) that of single network elements also has no bearing on these issues.
20. I think this characterization applies to contemporary philosophers as well, although I am not sufficiently well informed to be sure of this. Nagel appears to accept that there must be such a causal linkage, while doubting that we can make it intelligible. Colin McGinn (1999) has staked out a similar position which accepts the existence of causal connections but attempts to deny on principled grounds (unsuccessfully, I think) that we can ever understand them. Searle’s most radical philosophical opponent, however, is undoubtedly David Chalmers (1996). Chalmers argues that “zombies,” beings exactly like us behaviorally but completely unconscious, are logically conceivable, and that consciousness must therefore be something above and beyond the recognized physical order. Searle will have none of this, but his rebuttal again presumes what is in question, that low-level physical processes in the brain cause consciousness. See their exchange in Searle (1997).
21. Perhaps not coincidentally, detailed postmortem anatomical investigation of Einstein’s brain revealed that the most striking respect in which it differed from ordinary brains was in having significantly larger numbers of glia (Diamond, Schei-bel, & Harvey, 1985).
22. In Chapter 3 we begin to redress the asymmetry by describing numerous empirical phenomena of just this kind.
23. As indicated by an 1897 letter to Schiller (Perry, 1935, vol. 2, pp. 133–134), James thought initially that the transmission theory was his own invention, but it certainly has a much longer history. By the time of the Ingersoll lecture James himself had identified the following passage from Kant’s Critique of Pure Reason: “The body would thus be, not the cause of our thinking, but merely a condition restrictive thereof, and, although essential to our sensuous and animal consciousness, it may be regarded as an impeder of our pure spiritual life” (1898/1900, pp. 28–29). Michael Grosso informs me that the filter concept can also be detected in a number of the Platonic dialogues, including Phaedo, Phaedrus, and Ion. Clearly, a definitive history of the filter model is yet to be written.
24. A curious and relevant historical precedent is provided by Turing (1950), who explicitly considered the possibility that telepathy could undermine his proposed Turing-test procedures in favor of the human. Indeed, he evidently took this rather seriously, since it appears last in a list of objections ordered at least roughly in terms of increasing difficulty. His “solution” to the problem, however—putting the human into a “telepathy-proof room”—is patently defective, as he himself probably knew.
25. There is an important historical irony here. Dennett and Kinsbourne (1992) also focus on the absence of anatomical convergence, apparently thinking this is something new, but use it in a completely different way. Whereas McDougall (1911/1961) took the unity of conscious experience as a fundamental and undeniable empirical reality, one which physicalism could not explain, Kinsbourne and especially Dennett (1991) want to use the physiological diversity to undermine that appearance of unity itself, along with other supposedly pre-scientific “folk-psychology” intuitions about the nature of consciousness. See also Searle (1997, chap. 5), later sections of this chapter, and our Chapter 9.
26. This paragraph summarizes decades of cumulative progress in the neurobiology of sensory coding. The early work, inspired by Hubel and Wiesel (1962), emphasized “feature detection” by single sensory neurons, and provided for detection and representation of higher-order features or conjunctions of features by means of suitable anatomical connectivity. It was subsequently recognized, however, that combinations of elementary features could potentially occur in numbers far too large to manage exclusively in this anatomically-based way. The new functional proposals overcome the combinatorial explosion by providing for rapid and reversible linkages among groups of cells responding to more elementary properties.
27. Note that this characterization does not imply or require that the oscillatory activity itself satisfactorily explains binding. That it does not has already been argued by neurophysiologists such as Crick and Koch (2003) and Shadlen and Movshon (1999). See also the following paragraphs.
28. Although the “global workspace” terminology originated with Baars (1988, 1997), we will use it throughout this book in the more generic sense supplied in the text. Baars certainly deserves much credit for emphasizing that we need somehow to reconcile the unity of conscious experience with the multiplicity of associated neurophysiological processes in the brain, and for stimulating new imaging studies that seek to identify more precisely the critical neurophysiological conditions themselves (see, e.g., Dehaene & Naccache, 2001). His own specific version of a theory of this type, however, is less than satisfactory. In particular, his extended allegory of the “theater of consciousness” (Baars, 1997, pp. 41–47), although providing a colorful vocabulary with which to describe or interpret a variety of psychological phenomena, is conceptually incoherent in a multitude of ways.
29. Yuo mgiht aslo be srupsired to fnid taht yuo can raed tihs ntoe wtihuot mcuh truoble.
30. Another relevant phenomenon, and one that deserves more attention, is spontaneous hallucinatory experience in normal and awake persons. That such experiences commonly occur was initially demonstrated by the founders of the Society for Psychical Research (Gurney, Myers, & Podmore, 1886; H. Sidgwick et al., 1894) and has subsequently been confirmed repeatedly by others (Bentall, 2000). The mere fact of their occurrence demonstrates that the projective activity can sometimes partially or even wholly override current sensory input. Waking apparitions also share with dreams a tendency to incorporate massive amounts of information about physically remote events (Gurney et al., 1886). Critical readers should not indulge any temptation they may experience to dismiss these case reports wholesale as mere “anecdotes,” for they are massively and carefully documented. Penetrating analysis of the pitfalls of eyewitness testimony did not begin with Loftus (1979), as commonly supposed: See Gurney et al. (1886), its review by James (1887), and Chapters 2 and 6.
31. See also Chapter 4, which discusses several of these issues in the specific context of memory theory, and Gauld (1989), which develops views similar to those expressed here more systematically and generally in regard to the “entrapment” of cognitive science by its current conceptual framework.