Chapter 11 Reductionism and Antireductionism in Functionalist Theories of Mind Patricia Churchland

Antireductionism in Functionalist Theories of the Mind

Functional Types and Structural Implementations

The core idea of functionalism is the thesis that mental states are defined in terms of their abstract causal roles within the wider information-processing system.

In general, functional kinds are specified by reference to their roles or relational profiles, not by reference to the material structure in which they are instantiated. What makes a certain part of an engine a valve lifter is that, given a specified input, it has a certain output, namely the lifting of the valves, and it might be instantiated in various physical devices, such as a rotating camshaft or a hydraulic device. More humbly, "mousetrap" is a functional kind, being implementable in all manner of physically different devices: spring traps, assorted cage traps, a sack of grain falling when a trip line is wriggled, or perhaps even a cat or a specially bred killer rat. There is nothing in the specification "mousetrap" that says it must have a tin spring or a wooden housing.

According to functionalism, then, mental states and processes are functional kinds. Functionalists have typically sided with physicalism by claiming that our mental states are implemented in neural stuff, not, as the dualist would have it, in spiritual stuff. At one level of description we can talk about the causal and logical relations among perceptions, beliefs, desires, and behavior, and at the structural level we can talk about spiking frequencies of neurons, patterns of excitations, and so forth. It is because neurons are orchestrated as they are that the system has the functional organization it does, and thus the physical substratum subserves the functional superstratum. In our case the functional organization that is our psychology is realized in our neural "gubbins." In similar fashion, it is because on-off switches in a computer are orches­trated as they are that it adds, finds square roots, and so forth. The computer's program is realized in its electronic "gubbins." The functionalist theory is thus as roundly physi- calist as it can be, yet despite their adherence to physicalist principles, functionalists have typically rejected reductionism and ignored neuroscience. Why?

Plainly, it is not because functionalists suppose that mental states have no material realization. Rather, it is because they envision that types of mental states could have too many distinct material realizations for a reductive mold to fit. As functionalists see it, for a reductive strategy to succeed, a type of mental state must be identical to a type of physical state, but, they argue, the identities are not forthcoming. The reason is that one and the same cognitive organization might be realized or embodied in various ways in various stuffs, which entails that there cannot be one-to-one relations between func­tional types and structural types.

In a general way one can imagine that on another planet there might have evolved creatures who, though very different from us in physical structure, might have a cogni­tive organization much like our own. Suppose, for example, they were silicon-based instead of carbon-based as we are. For these animals, having a goal will be functionally like our having a goal, but such a state will not be identical to having neurons n-m responding thus and so, though to be sure the goal state will be embodied in their physical structure. Or suppose that in time we figure out how to manufacture a robot that has the same functional organization as a human: it has goals, beliefs, and pains, and it solves problems, sees, and moves about. Its information-processing innards are not neurons but microchips, and its cognitive organization cannot therefore be identical to a particular neuronal organization, since neural stuff it has not got. Instead, its cognitive economy will be instantiated in electronic stuff. As we shall see, the plausibility of these thought-experiments depends on a crucial and highly suspect assumption—namely, that we know at what level the biology does not matter.

Fictional examples are not really needed to make the point anyhow, since there are certain to be neural (structural) differences between functionally identical states in dis­tinct species.

Identity of functional-state types with structural-state types, argues the functionalist, is therefore hopelessly unrealistic, and since reduction requires such identities, tant pis

Reductionism and Antireductionism in Functionalist Theories of Mind 61 for reduction. Physicalist principles are in no way sundered, however, for all that physicalism requires is that any given instance of a functional-state type (a token of that type) be realized in physical stuff, and this the functionalist heartily agrees to and insists upon. He therefore describes himself as espousing token-token identity of mental states with physical states, but denying type-type identity and therewith reductionism (Putnam 1967, Dennett 1978b).

This foray against the reductionist program is known as the argument from multiple Instantiability or multiple realizability. Functional states are multiply instantiable, and the range of physical implementations will be so diverse that we cannot expect it to form a natural kind.

If mental states and processes are functional kinds, then to understand how cogni­tively adept organisms solve problems, think, reason, and comport themselves intelli­gently, what we need to understand is their functional organization. Research on neurons is not going to reveal the nature of the functional organization, but only something about the embodiment of the functional organization—and just one sort of instantiation at that. Neuroscience, it has been argued, is focused on the engineering details rather than on the functional scheme, and to this extent it is removed from the level of description that is appropriate to answering questions concerning learning, intelligence, problem solving, memory, and so forth. Knowledge of the structural minu­tiae is important for repairs, of course, and to this extent neuroscience has obvious medical significance, but structural theory will not enlighten functional hypotheses and functional models. To put it crudely, it will not tell us how the mind works. Cognitive psychology, in contrast, is focused at the appropriate level of description, and in coop­eration with research in artificial intelligence it constitutes the best strategy for devising a theory of our functional cognitive economy. Thus the crux of the argument.

As Pylyshyn (1980) sees it, the research labor can be divided along these lines: the cognitive scientists will figure out the functional/cognitive theory, and the neuro­scientists can untangle the underlying physical devices that instantiate the cognitive "program." On an extreme version of this view, nothing much of the details of neuronal business need be known by the cognitive scientist—or the philosopher, either—since the way the functional Organization is instantiated in the brain is a quite separate and independent matter from the way our cognitive economy is organized.

Neuroscience, on this picture, is irrelevant to the computational questions of cogni­tive science. What it is relevant to are implementation issues, such as whether a particu­lar computational model of cognitive business is in fact implemented in the neural structure. Computational (functional) psychology is thus conceived as an autonomous science, with its proprietary vocabulary and its own domain of questions, the answers to which, as Pylyshyn remarks, "... can be given without regard to the material or hardware properties of the device on which these processes are executed" (p. 115). It may even be suggested that the less known about the actual pumps and pulleys of the embodiment of mental life, the better, for the less there is to clutter up one's function­ally oriented research.

Whether anyone really holds the extreme version of the research ideology is doubt­ful, but certainly milder versions have won considerable sympathy, and sometimes cognitive science programs permit or encourage neglect of neuroscience, where the autonomy of psychology is the rationale. How influential the view is I cannot estimate, but some philosophers are still wont to excuse those colleagues who take neuroscience seriously as having not quite managed to master the distinction between functional and structural descriptions. The methodological point should be taken seriously because functionalism is now the dominant theory of the mind espoused by philosophers as well as by many cognitive scientists. Even so, there are significant differences among functionalists on a number of issues, including the relevance of theories of brain func­tion to theories of psychological function. Dissent from the methodological point is not without voice in cognitive psychology (for example, McClelland and Rumelhart 1981, Posner, Pea, and Volpe 1982), philosophy (for example, Enς 1983, Hooker 1981, Paul M. Churchland 1981), and computer science (for example, Anderson and Hinton 1981). My lot is thrown in with the dissenters, because I think both the antireductionist argument and the research ideology it funds are theoretically unjustified and pragmati­cally unwise to boot. In what follows I shall try to show why.

In Defense OfReductionism

There are two principal sources of error in the antireductionist views I have outlined. The first concerns the background assumptions about the nature of intertheoretic reduc­tion; the second concerns the conception of levels—how many there are, their nature, their discovery, and their interconnections. These sources of error will be considered in sequence.

Intertheoretic Reduction and Functionalism

Functionalists appear to assume that intertheoretic reduction cannot come off unless the properties in the reduced theory have a unique realization in physical stuff. This assump­tion is crucial in the case against reduction, and it is what floats the methodological claim for the autonomy of cognitive psychology. Is the assumption justified?

One way to test the claim is to see whether it conflicts with or comports with the paradigm cases of reduction in the history of science. "Temperature" is a predicate of thermodynamics, and as thermodynamics and molecular theory co-evolved, the tem­perature of gases was found to reduce to the mean kinetic energy of the constituent molecules. That is, a corrected version of the classical ideal gas law was derived from statistical mechanics together with certain assumptions. Several features of this case are immediately relevant to the issue at hand. Notice that what was reduced was not temperature tout court, but temperature of a gas. The temperature of a gas is mean kinetic energy of the constituent molecules, but the temperature of a solid is something else again; the temperature of a plasma cannot be a matter of kinetic energy of the molecules, because plasmas are high-energy states consisting not of molecules but of dissociated atoms; the temperature of empty space as embodied in its transient electro­magnetic radiation is different yet again. (Paul M. Churchland 1984 and Enς 1983 also make this point.) And perhaps there are states as yet undiscovered for which tempera­ture is specified in none of these ways. The initial reduction in thermodynamics was relative to a certain domain of phenomena, to wit, gases, but it was a bona fide reduc­tion for all that. Nor is this domain-relativity used as grounds for saying that thermody­namics is an autonomous science, independent and separate from physics. Quite the

Reductionism and Antireductionism in Functionalist Theories of Mind 63 contrary, the co-evolution of corpuscular physics and thermodynamics was of the first importance to both physics and thermodynamics.

Yet if we heed the functionalist assumption at issue, we ought to withhold the stamp of reduction on grounds that temperature must be a functional property that is multiply realized in distinct physical structures. Now, however, this looks like a merely verbal recommendation about what to call reductions in cases where the predicates in the reducing theory are relativized to certain domains (cf. Cummins 1983). As such, it implies nothing about the derivation of one theory from another or about the autono­my of the sciences. No grand methodological strictures about what is and is not rele­vant to the "functional" theory will be in order. As a merely verbal recommendation it is not especially objectionable, but it has no obvious utility either..

Dialectically, it does the functionalist no good to deny reduction in thermodynamics, for then he loses the basis for saying that psychology is on an entirely different footing from the rest of science (Enς 1983). After all, if psychology is no worse off than thermodynamics, then reductionists can be cheerful indeed. At any rate, the require­ments for the reduction of psychology should not be made stiffer than those for intertheoretic reduction elsewhere in science. (See also Richardson 1979, Paul M. Churchland 1984.)

The main point of the example drawn from thermodynamics is that reductions may be reductions relative to a domain of phenomena. Though this is called "multiple instanti- ability" and is draped in black by the functionalist, it is seen as part of normal business in the rest of science. By analogy with the thermodynamics example, if human brains and electronic brains both enjoy a certain type of cognitive organization, we may get two distinct, domain-relative reductions. Or we may, in the fullness of time and after much co-evolution in theories, have one reductive account of, say, goals or pain in vertebrates, a different account for invertebrates, and so forth. In and of itself, the mere fact that there are differences in hardware has no implications whatever for whether the psychology of humans will eventually be explained in neuroscientific terms, whether the construction of psychological theories can benefit from neuro­scientific information, and whether psychology is an autonomous and independent science. That reductions are domain-relative does not mean they are phony reductions or reductions manque, and it certainly does not mean that psychology can justify methodological isolation from neuroscience.

Enς (1983) draws a further point out of the thermodynamics case. Two volumes of a gas might have the same temperature, but the distributions of velocities of their constituent molecules will be quite different even while their mean value is the same. To be consistent, functionalists should again deny reductive success to statistical mechanics since, as they would put it, temperature of a gas is differently realized in the two cases. If, on the other hand, they want to concede reduction here but withhold its possibility from psychology, they need to do more than merely predict hardware differences between species or between individuals.

If it turns out that we are lucky enough to get a reduction (domain-relative) of human psychology to neuroscience, what does this do to the thesis that mental kinds are functional kinds? Nothing, for that thesis is independent of the antireductionist argu­ment, and it stands on its own feet after the argument from multiple instantiability falls. The thesis that mental states are identified in terms of their abstract causal roles in the wider information-processing system is the core conception that makes functionalism functionalism, and it is entirely neutral on the question of reducibility. Functionalists can be true blue functionalists without naysaying reduction. Functionalism as it lives and breathes, however, is another matter, and frequently functionalists have wished to argue for a package: the functional characterization of mental states, the nonredudbility of psychology, and the autonomy (in some degree) of psychology from the more basic sdences. As a result, the term "functionalism" is typically if inappropriately assodated with the whole package.

The point of this section has been a very general one: intertheoretic reductions are not conditional on a one-to-one mapping of predicates of the higher-level theory onto predicates of the reducing theory. Antireductionists may wish to concede the general point but to continue by arguing that the details of the case at hand rule out reduction. In so arguing, they will point to radical differences between the neuronal level of explanation and the functional-computational level, and they will point out that the Inultiplidty of instantiations of psychological predicates can be so profuse, diverse, and arbitrary that the case cannot be likened to the thermodynamics-statistical mechanics example. In a word, they daim that the case of psychology is spedal.

Levels of Organization in the Mind-Brain

There is a good deal that is Uncontroversial in the antireductionist's appredation that there must be a set of levels of organization. A theory of cellular and synaptic changes occurring during learning will be more fine-grained than a theory of how an interactive network learns, which will be more fine-grained than a theory of what anatomical structures subserve learning, which will be more fine-grained than a theory that postu­lates a coding mechanism, retrieval mechanisms, and so forth. What is controversial is the assumption that the trilevel model suitable to Von Neumann computers is also suitable to organic brains. That there should be some division of labor is also beyond dispute; no one since Bacon could take all knowledge as his province. Indeed, no one since Helmholtz could take even all of neuroscience as his province. What is regrettable, however, is the divisive research ideology based on the trilevel model.

As we have seen, the hypothesis based on the computer analogy is that the mind­brain has three levels of organization: the semantic, the syntactic, and the mechanistic— the level of content, the level of the algorithm, and the level of structural implementation. The principal problem with the computer metaphor is that on the basis of the com­plexity we already know to be found in the brain, it is evident that there are many levels of organization between the topmost level and the level of intracellular dynamics. (See also Lycan 1981a.) And even if there were just three, neurobiological theory challenges that way of specifying their organizational description. How many levels there are, and how they should be described, is not something to be decided in advance of empirical theory. Pretheoretically, we have only rough and ready—and eminently revisable— hunches about what constitutes a level of organization.

As a first approximation, we can distinguish the following levels of organization: the membrane, the cell, the synapse, the cell assembly, the circuit, the behavior. And within each level further substrata can be distinguished. If, however, neurons are organized into modules, each perhaps playing a role in several distinct information-processing modules, and if modules themselves are members of higher-order "metamodules," again with membership being a diverse and distributed affair, or if some cell assemblies or modules have a transient membership or a transient existence, we may then find a description of levels that is orthogonal to the first.

Another preliminary and related way to demarcate a level is to characterize it in terms of the research methods used. Certainly this is a very rough way of defining levels of organization, but it may be useful until the research reveals enough for us to see what the levels really are. For example, in research on learning and memory one can discern many different methods that, compared to one another, are more or less fine-

Reductionism and Antireductionism in Functionalist Theories of Mind 65 grained. The cellular approach taken by Kandel and his colleagues (Hawkins and Kandel 1984) showing modification in presynaptic neurotransmitter release in habituation is in some sense at a lower level than studies by Lynch and his colleagues (Lee et al. 1980) showing modification of synapse numbers and synaptic morphology correlated with plasticity in behavior, which in turn is at a (slightly) lower level than the studies by Greenough and his colleagues (Greenough, Juraska, and Volkmar 1979) on the effect of maze training on dendritic branching. We then ascend to the multicellular studies in the hippocampus done by Berger, Latham, and Thompson (1980), and from there up (a bit) to the cell assembly studies in the olfactory bulb by Freeman (1979), which uses an 8 ? 8 electrode array and evoked response potential averaging techniques. Upward again to the studies of Nottebohm (1981) on the seasonal changes in the "songster" nuclei of the canary brain or to the animal models of human amnesia studied by Zola-Morgan and Squire (1984). At yet a different level are the studies by Jemigan (1984) and Volpe et al. (1983) of correlations between neural tissue atrophy and memory performance using neural imaging techniques (CBF, PET). Finally there are neurological studies of human amnesia (Weiskrantz 1978, Squire and Cohen 1984), ethological studies of such things as how bees remember flowers (Gould 1985), and psychological studies of memory capacities and skills of college undergraduates (Norman 1973, Tulving 1983). This is obviously a very fast Cook's ascent at just one point through the research strata, but a more leisurely tour will reinforce the impressions.

It is simply not rewarding to sort out this research in terms of the trilevel computer analogy, nor is there any useful purpose to be served by trying to force a fit. Moreover, at each of the research levels one can distinguish among questions concerning the nature of the capacity, questions concerning the processes subserving the capacity, and the matter of the physical implementation. The point is, even at the level of cellular research, one can view the cell as being the functional unit with a certain input-output profile, as having a specifiable dynamics, and as having a structural implementation in certain proteins and other Subcellular structures.

What this means is that one cannot foist on the brain a monolithic distinction between function and structure, and then appoint psychologists to attend to function and neuroscientists to attend to structure. Relative to a lower research level a neurosdentist's research can be considered functional, and relative to a higher level it can be considered structural. Thus, Thompson's work on multicellular response profiles in the hippocampus is perhaps structural relative to Squire's work on the recognition capacities of amnesic humans but functional relative to Lynch's work on plasticity of synaptic morphology. The structure-function distinction, though not without utility, is a relative, not an absolute, distinction, and even then it is insufficiently precise to support any sweeping research ideology.

In addition, we simply do not know at what level of organization one can assume that the physical implementation can vary but the capacities will remain the same. In brief, it may be that if we had a complete cognitive neurobiology we would find that to build a computer with the same capacities as the human brain, we had to use as structural elements things that behaved very like neurons. That is, the artificial units would have to have both action potentials and graded potentials, and a full repertoire of synaptic modifiability, dendritic growth, and so forth, though unlike neurons they might not need to have, say, mitochondria or ribosomes. But, for all we know now, to mimic nervous plasticity efficiently, we might have to mimic very closely even certain Subcellular structures.

There is a further assumption, usually unstated, that lends credence to the ideology of autonomy and should be debunked. This assumption is that neuroscience, because it tries to understand the physical device—the brain itself—will not produce theories of functional organization. Now we have already seen that the functional-structural dis­tinction will not support the simplistic idea that psychology does functional analysis and neuroscience does structural analysis, and that there are bound to be many levels of organization between the level of the single cell and the level at which most cognitive psychologists work. It is important as well to emphasize that when neuroscientists do address such questions as how neurons manage to store information, or how cell assemblies do pattern recognition, or how they manage to effect sensorimotor control, they are addressing questions concerning neurodynamics—concerning information and how the brain processes it. In doing so, they are up to their ears in theorizing, and even more shocking, in theorizing about representations and computations. If the rep­resentations postulated are not sentencelike, and if the transformations postulated do not resemble reasoning, this does not mean the theory is not functional theory, or not real theory, or not relevant to theories at a higher level.... The existence of bona fide neurofunctional theorizing is perhaps the most resounding refutation of the second assumption.

My general conclusion, therefore, is that it is supremely naive to assume that we know what level is functional and what is structural, and that neurons can be ignored as we get on with the functional specification of the mind-brain. This explains my earlier warning about the multiple instantiation thought-experiments that are endlessly in­voked by antireductionists. Nevertheless, antireductionists will argue for the autonomy of cognitive psychology not merely on the basis of the trilevel hypothesis but also on the grounds that the categories and generalizations appropriate to the cognitive levels are special. For reasons to be examined, these categories are believed to have an invulnerable theoretical integrity and to be irreducible to physical categories.

Bibliography

Anderson, James A. and Geoffrey Hinton, 1981. "Models of Information Processing in the Brain." In Hinton and Anderson, eds., Parallel Models of Associative Memory. Hillsdale NJ: Erlbaum.

Berger, T. W., R. I. Latham, and R. F. Thomson, 1980. "Hippocampal Unit Behavior Correlations During Classical Conditioning," Brain Research 193:483-85.

Churchland, Paul M., 1981. "Eliminative Materialism and the Propositional Attitudes." Journal of Philosophy 78, no. 2,67-90.

Churchland, Paul M., 1984. Matterand Consciousness: A Contemporary Introduction to the Philosophy of Mind. Cambridge MA: MIT Press.

Cummins, Robert, 1983. The Nature of Psychological Explanation. Cambridge MA: MIT Press.

Dennett, Daniel C., 1978b. Brainstorms: Philosophical Essays on Mind and Psychology. Cambridge MA: MTT Press.

Enς, Berent, 1983. "In Defense of the Identity Theory," Journal of Philosophy 80:279-98.

Fodor, Jerry A., 1975. The Language of Thought. New Yoric Crowell.

Freeman, Walter, 1979. "Nonlinear dynamics of paleocortex manifested in the olfactory EEG," Biological Cybernetics 3: 21-37.

Gould, James L., 1985. "How Bees Remember Flower Shapes," Science 227:1492-94.

Greenough, W. T., J. M. Juraska, and F. R. Volkmar, 1979. "Maze Training Effects on Dendritic Branching in Occipital Cortex of Adult Rats," Behavioral and Neural Biology 26:287-97.

Hawkins, Robert D., and Eric R. Kandel, 1984. "Steps Toward a Cell-Biological Alphabet for Elementary Forms of Learning." In G. Lynch J. L. McGaugh and N. M. Weinberger, eds., Neurobiology of Learning and Memory. New York: Guilford.

Hooker, Clifford A., 1981. 'Toward a General Theory of Reduction," Dialogue 20: 38-59, 201-236, 496-529.

Jemigaa Terry L, 1984. "The Study of Human Memory with Neuro-Imaging Techniques." In L R. Squire and N. Butters, eds., Neuropsychology of Memory. New Yoric Guilfdrd.

Lee, K. S., F. Schottler, M. Oliver, and G. Lyndι, 1980. "Brief Bursts of High-Frequency Produce Two Types of Structural Change in Rat Hippocampus," Journal PfNeurophysiology 44:247-58.

Lycaa William G., 1981a. "Form, Functioa and FeeL" Journal of Philosophy 78:24-50.

Lycaa William G., 1981b. 'Toward a Homuncular Theory of Believing," Cognition and Brain Theory 4: 139-59.

McOelland, James L, and David E. Rumelhart, 1981. "An Interactive Activation Model of the Effect of Context in Letter Perceptioa Part I: An Account of Basic Findings," Psychological Review 88:375-407.

Normaa D. A, 1973. 'Memory, Knowledge, and the Answering of Questions." In R. Solso, ed., The Loyola Symposium on Cognitive Psychology. Washington D.C.: Winston.

Nottebohm, F., 1981. "Laterality, Seasons, and Space Governing the Learning of a Motor SkiU," Trends in Neuroscience 4, no. 5:104—6. *.

Posner, Michael L, Roy Pea, and Bruce Volpe, 1982. "Cognitive Neurosdence: Developments Toward a Sdence of Synthesis." In J. Mehler, E. Walker, and M. Garrett, eds., Perspectives on Mental Representa­tion. Hillsdale NJ: Eribaum.

Putnam, Hilary, 1967. "The Nature of Mental States." In W. H. Capitan and D. D. MerriU, eds., Art, Mind, and Religion. Pittsburgh: Univ. Pittsburgh Press.

Pylyshya Zenoa 1980. "Computation and Cognition: Issues in the Foundation of Cognitive Sdence," Behavioral and Brain Sciences 3, no. 1:111-34.

Pylyshya Zenoa 1984. Computation and Cognition. Cambridge MA: MIT Press.

Ridiardsoa Robert, 1979. 'Tunctionalism and Reductioa" Philosophy of Science 46:533-58.

Squire, Larry R., and Neal J. Cohea 1984. "Human Memory and Amnesia." In G. Lynch, J. L McGaugh, and N. M. Weinberger, eds., Neurobiology of Learning and Memory. New York: Guilford.

Tulving, EndeL 1983. Elements of Episodic Memory. Oxford: Clarendon Press.

Volpe, B. T., P. Herscovitch, M. E. Raichle, M. S. Gazzaniga, and W. Hirst, 1983. "Cerebral Blood Flow and Metabolism in Human Amnesia," Journal of Cerebral Blood Flow and Metabolism 3:5.

Weiskrantz, L, 1978. "A Comparison of Hippocampal Pathology in Man and Other Animals." In Functions of the Septo-Hippocampal System. Amsterdam: Elsevier.

Zola-Morgaa S., and L R. Squire, 1984. "Preserved Learning in Monkeys with Medial Temporal Lesions: Sparing of Motor and Cognitive Skills," Journal of Neuroscience 4:1072-85.