The theory of multiple intelligences: A case of missing cognitive matter.

Introduction

Nearly 20 years ago, Howard Gardner mounted a challenge to conventional thinking on the nature of human intelligence with the presentation of a provocative new conception, which was neatly encapsulated in the title of his 1983 book, Frames of mind: The theory of multiple intelligences. In this work, Gardner mapped out an alternative to the prevailing orthodoxy that intelligence in humans is a general intellectual ability or global factor, which permeates all aspects of cognition, and most effectively predicts an individual's performance in school or work contexts, and in many other aspects of well-being in life. As such, this general factor, known as g, which is extracted statistically by means of a technique called factor analysis, is for those involved in mainstream intelligence research and psychometric testing, the working definition of intelligence, and is interchangeable with the more common expression of IQ (Gottfredson, 1998, pp. 24-25).

Although not disputing the existence of g, Gardner nevertheless argues that the conventional view of a single general intelligence is far too narrow to be a useful measure of human potential and capability, for it fails to come to grips with higher capacities such as creativity, and is insensitive to a range of socio-cultural roles (Gardner, 1993a, pp. 24, 39). In particular, he charges that the IQ view of intelligence is consistent with the concerns of traditional psychology, and is a `one-dimensional' view of mental ability that focuses principally on logical-mathematical and linguistic reasoning, both of which have corresponded with a conventional view of schooling and mental assessment (see Gardner, 1993b, pp. 5-6, 20-21). In contrast to this view, and as a result of his own work in developmental psychology, Gardner concluded that human intelligence encompasses a much broader, and more universal set of competencies which, as relatively independent faculties of mind, constitute a collection of discrete and more or less autonomous intelligences. These include the traditional logical-mathematical and linguistic forms of intelligence, and some less conventional forms such as musical, spatial, bodily kinesthetic, interpersonal, and intrapersonal intelligences (Gardner, 1993a, pp. 73-78, 1993b, pp. 17-26, 1998). In all, Gardner originally identified seven distinct intelligences, but later added an eighth `naturalist' form of intelligence in 1995, and more recently has been considering a ninth form of `existential' intelligence (Gardner, 1998, p. 21). Gardner's alternative therefore incorporates a pluralistic view of mind and intellect, which recognises many different and discrete facets of cognition, and possible profiles of mental capability and competence. Instead of there being a single dimension of intellect along which individuals can be rank-ordered, there are instead considerable differences in the intellectual strengths and weaknesses of various individuals, and in the ways in which they approach cognitive tasks (Gardner, 1993a, pp. 6, 169-170, 1994, p. 740).

In what follows, consideration will be given to some of the central features of Gardner's conception, where attention will focus in particular on a methodological evaluation of the theory's content and structure from within a naturalistic-coherentist framework for the justification of knowledge, in an attempt to discern how well this alternative to the mainstream stands or falls in terms of the explanatory tasks that it sets for itself.

Methodological considerations

Although the theory of multiple intelligences (MI theory) has generated controversy in psychological quarters, it has nevertheless been more generously received in educational circles, as it has accorded well with educator intuitions that children are intelligent in different ways, and that different educational approaches may reach students more effectively, if favoured ways of learning and knowing are accounted for in the areas of curriculum, instruction, and assessment (Gardner, 1998, p. 21). More specifically, MI theory is seen to offer educators a potentially more viable account of cognitive functioning, as it is based on, and takes into account, the nature of real-world intelligent behaviour, such as social and interpersonal skill, and the creative domains of artistic talent or giftedness. Furthermore, in providing a framework for discriminating between different individual intellectual profiles, MI theory is also seen to be of corresponding practical value to educators, in that it facilitates the identification of particular or special learning needs, and the programming of appropriate educational responses to them. Nevertheless, despite these perceived virtues, overall support for MI theory in cognitive science research literature appears to be inconclusive and rather mixed. In particular, there is `little if anything in the way of hard evidence ... for Gardner's theory' and efforts to validate it empirically have proved to be `very difficult, if not impossible' (see Matthews, 1995, pp. 229-244).

However, assessing the virtues of a conceptual framework such as MI theory requires more than hard evidence, or mere empirical adequacy alone, for recent developments in the philosophy of science have shown that the epistemological basis for theoretical virtue in any domain requires considerably more than this, and extends to consideration of a much wider set of super-empirical epistemic criteria that are associated with the global and systemic virtues that theories possess, such as relative coherence, consistency, comprehensiveness, simplicity, explanatory unity, and learnability, which together constitute a coherentist framework for the justification of knowledge (see BonJour, 1985; Churchland, P.M., 1989, pp. 139-151; Evers & Lakomski, 1991, pp. 1-45; Quine & Ullian, 1970, pp. 42-53; Williams, 1977, 1980, pp. 243-272). Analysis of empiricist doctrine has shown that theories are underdetermined by available evidence, and that observation statements on their own convey very little about the nature of the empirical world, as they are always embedded in a much wider network of informing statements, many of which have no direct connection with sensory perception. Thus the most basic units of meaning are whole theories in themselves, and it is these that have empirical consequences. Hence, as theory-laden wholes, observations are not epistemically privileged or foundational, as positivist and empiricist accounts of knowledge justification have presumed (see Churchland, P.S. 1992, pp. 265-271; Quine, 1951, pp. 20-43).

Given this methodological development, it is not surprising, therefore, that efforts to validate or, conversely, to disconfirm the theory of multiple intelligences have been problematic, for the resources of empiricism are too meagre for this task. Although Gardner has attempted to produce a scientific theory, and one that makes a contribution to cognitive science (see Gardner, 1993a, pp. xxvi, 10), he nevertheless gives little, if any, concerted attention to matters of epistemology and its bearing on methodological issues in theory construction. This is a rather curious incongruity, considering that, in a self-referential sense, MI theory deals with fundamental questions concerning the nature of mind and cognition. Nonetheless the importance of a coherent theory of mind for methodological and theoretical developments can be seen when consideration is given to what an epistemology claims as being sufficient or adequate for the justification of knowledge, for this functions to constrain the substantive content and structure of theorising, and plays a central role in shaping the nature and character of theoretical, methodological, and practical developments in science.

In specifying conditions for knowledge claims to count as justified, an empirically plausible theory of perception, learning, memory, representation, and cognition is essential. If there is no first philosophy or incorrigible foundation for knowledge justification, as required by empiricist theories of knowledge, then as a corrigible theory of mind and knowledge in its own right, epistemology falls into place within the wider fabric of scientific knowledge as a chapter of psychology, and once naturalised in this way, our best theories of learning and cognition are then used to explain and justify how scientific knowledge is possible (see Evers, 1988, p. 10; Evers & Lakomski, 1991, p. 8; Quine, 1969, p.82). Thus learnability constitutes a self-referential constraint on any theory of knowledge, in that it must cohere with what our best existing theories of mind and cognition have to say about how we come to learn about our world, and correspondingly how we can come to know the nature of our own minds. Consequently, for this and other reasons to be outlined below, a naturalistic and coherentist conception of knowledge justification is to be preferred in the task of evaluating how well MI theory stands or falls as an account that examines one of the more significant aspects of mind, namely intelligent cognition.

Gardner's conception, and the frame of mind that informs it

Since methodological issues are crucial to the task of theory evaluation, it will be instructive to begin by noting explicitly what Gardner has to say about the scientific method, for his views disclose the background theory of knowledge underwriting his conception. On this matter, Gardner (1993a) writes:

   In the scientific method, at least two aspects work hand in hand. On the
   one side, there is an interest in collecting facts, a desire to be
   objectively empirical and to find out as much as one can about a subject
   matter, along with a willingness ... to change one's mind in the light of
   new facts. Complementing this descriptive aspect of science is the building
   of an explanatory superstructure--a theoretical framework which explains
   the nature of, and the relations among, objects and forces, how they come
   about, what can make them change, and under what condition such changes are
   likely to take place. The theoretical structure depends upon reasoning:
   deductive reasoning, where implications are drawn from general assumptions;
   and inductive reasoning, where general principles are arrived at from the
   examination of individual cases. (p.361)

As it stands, this characterisation of science accords with that of logical empiricism, which as an epistemology is no longer accepted as valid in the philosophy of science, and in neurophilosophical inquiry into knowledge representation and computation in the mind/brain (see Churchland, P.M., 1989; Churchland, P.S., 1987, 1992; Evers & Lakomski, 1991, pp. 1-45; Quine 1951, 1969). Its key feature is that knowledge consists of a set of claims that are organised into a hypothetico-deductive structure or theory, which has the following main elements:

1 Theories consist of a set of topic-specific, empirical claims that are arranged in a hierarchy, where particular claims, such as singular observation statements about some phenomenon, are deduced from more general claims, including claims about the context in question.

2 Theories are justified through a process of empirical testing where, if the deduced lower-level claims or hypotheses in the hierarchy are actually observed in the testing process, then the theory is considered confirmed, or otherwise disconfirmed if contrary observations are made. In essence, the more confirmations a theory has vis-a-vis disconfirmations, the more justified the theory appears to be.

3 All terms in a theory must admit of some defining measurement procedure, or empirical definition that refers to some set of observations. This requirement presents difficulties in the case of so-called theoretical terms which, unlike observational terms, refer to unobservables. However, this difficulty is met by giving theoretical terms operational definitions, which amount to specifying some set of corresponding operations performed in measurement or application in relation to observable objects.

4 Since theoretical terms or statements acquire their meaning either directly from observations, or derivatively via operational definitions, those satisfying neither condition, such as moral, ethical, or aesthetic terms or statements, are considered meaningless, despite any meaningful content they might otherwise appear to have, since they are mere expressions of subjective feeling that cannot be known, and consequently are to be treated as opinion, attitude, or emotive utterance. Thus an epistemological cleavage exists between matters of fact, which are considered to be empirical and objective in content, and matters of value, which are deemed to be empirically unverifiable, and are therefore no more than subjectively motivated theoretical claims (see Ayer, 1975, p. 29; Evers, 1995, p. 2; Evers & Lakomski, 1991, pp. 1-45, 1996, pp. 383-384).

Logical empiricism therefore depends on maintaining a sharp distinction between theoretical and observational terms, and between facts and values. These epistemological features can be seen in the content and structure of MI theory, and this has implications for the theory's overall coherence and explanatory power.

Gardner first arrived at his intuitions concerning the multi-faceted nature of human intelligence from a synthesis of a number of sources of data, especially evidence from neurological, evolutionary, and cross-cultural origins, which had not been considered together before. Included was information about the development of different kinds of skills in normal children, information on the ways in which these abilities break down in cases of brain damage, information from special populations such as idiot savants, autistic children, and children with learning disabilities, evidence from experimental findings and psychometric correlations, cognition in diverse animal species, and anthropological studies. From these new sources of knowledge, Gardner saw the need for a better classification of human capabilities, and description was seen to be an appropriate place to begin this task. Although some sort of algorithm for optimally sorting out the details or factors from the mass of available data would have been desirable, the unavailability of such a procedure meant that Gardner had to resort more to an artistic judgement than a scientific assessment, by performing a kind of `subjective' factor analysis, to arrive at his map of the intelligences (see Gardner, 1993a, pp. xxv-xxvi, 1993b, pp. 7-8, 1998, p. 20). Hence he asserts that `the best way to start to understand the human mind is to examine its different frames, its separate intelligences'. Since the various intelligences usually work in harmony in ordinary life, their discreteness and autonomy may be invisible. However, `when the appropriate observational lenses are donned', the nature of each emerges with sufficient clarity (see Gardner, 1993a, pp. xxix, 9). To facilitate identification, Gardner developed a set of criteria that each posited intelligence had to meet if it was to count as a legitimate category. These included: potential isolation by brain damage; the existence of relatively isolated but exceptional abilities in certain individuals, such as idiot savants and prodigies, etc.; an identifiable core operation or set of operations; a distinctive developmental history within an individual, along with a recognisable expertness in performance; an evolutionary history and plausibility; experimental support from psychological tests; support from psychometric findings; and susceptibility to encoding in a symbol system (see Gardner, 1993a, pp. 62-66, 1998, p. 20).

Of particular importance to the task of delineating the separate intelligences is biological and neurological evidence. The pattern of breakdown in the behavioural and cognitive abilities of brain-damaged persons, and neuroscientific evidence for the existence of distinct and localised functional units or special-purpose information-processing devices in the brain, suggests that a biological basis underpins the specialised intelligences. Central to this conclusion, is that specific kinds of informational input are processed by these computational mechanisms, and that, as such, each form of intelligence is identified by the core operation, or set of operations that it performs. Hence each intelligence is `triggered' or activated by the presentation of certain kinds of internal or external information. Thus a core component of musical intelligence is sensitivity to pitch relations in auditory input, whereas phonological features constitute a core component of linguistic intelligence, for instance. Furthermore, as distinct systems, each intelligence operates according to its own procedures, with its own rules. Consequently neurobiological research hints strongly that the various forms of human intelligences constitute `natural kinds' (see Gardner, 1993a, pp. 36-58, 63-70, 1993b, p. 16).

However the intelligences cannot be seen solely as raw biological potentials, for in their pure form they are only useful scientific `fictions', or constructs, and as such are impossible to measure. Rather `intelligences are always expressed in the context of specific tasks, domains, and disciplines', and can only be assessed in such situations (Gardner, 1993a, pp. xx, 69). According to Gardner, this shift in assessment philosophy, and the distinctions drawn between intelligences, domains, and fields, together constitute possibly the most important conceptual advance in MI theory; for socialisation and the achievement of competence in various cultural domains can only be assessed by the field of persons, institutions, and other social structures that make up the area of capability in question, and which in turn are eligible to render judgements about the qualities of individual performance in those contexts. Thus, although there may be a biological basis for `natural kinds' of intelligences, culture cannot be neatly filtered out of the equation, for the development and implementation of intellectual competencies is only possible in cultural contexts. Consequently Gardner's conception of intelligent cognition arises from efforts to synthesise insights that have been gleaned from research in biology and culture (see Gardner, 1993a, pp. 57-58).

When it comes to linking raw biological potential with cultural content, Gardner adheres to the symbol systems approach in cognitive science, for `symbols pave the royal route from raw intelligences to finished cultures' (Gardner, 1993a, p. 302). Operationally defined as `culturally derived systems of meaning which capture important forms of information', symbols are seen to be ideal as entities, and indispensable as a level of analysis, for spanning the gap between biology and anthropology. Indeed `one of the features that makes a raw computational capacity useful ... is its susceptibility to marshalling by a cultural symbol system' (see Gardner, 1993a, pp. 66, 302). Hence `an intelligence must ... be susceptible to encoding in a symbol system' (Gardner, 1993b, p. 16). Thus one of the principal missions of modern education may be seen as involving the introduction and mastery of various symbol systems (Gardner, 1993a, p. 304). Consequently Gardner's research agenda at Harvard Project Zero has been to define and delineate particular symbolic domains by searching for and arriving at the `natural kinds' of symbol systems that either hang together or fall apart, and to discover how these may be represented in the human nervous system (Gardner, 1993a, pp. 28-29).

An intelligence is defined as `a psychobiological potential to solve problems or to fashion products that are valued in at least one cultural context' (Gardner, 1998, p. 20). In this light, human cognitive performance is more adequately described as consisting of a set of culturally significant abilities, talents, or mental skills which, as distinct competencies satisfying the prerequisite criteria of identification and delineation, constitute the so-called `intelligences' (see Gardner, 1993b, p. 15). Thus the sets of intelligences that have been included in MI theory are ones that have met certain biological and psychological specifications (see Gardner, 1993b, p. 61). Consequently `a prerequisite for a theory of multiple intelligences, as a whole, is that it captures a reasonably complete gamut of the kinds of abilities valued by human cultures' (Gardner, 1993a, p. 62). However the ultimate arbiter that underwrites this account of cognition comes from `a deep knowledge of the nervous system', and from `findings from the brain' (Gardner, 1993a, p. 30).

Problems with MI theory

Given the heavy demands that logical empiricism makes on available epistemological resources when constructing theory, and given the great complexity of neuropsychological and socio-cultural phenomena (see Evers & Lakomski, 1993, pp. 141-142), it is not surprising to find that MI theory runs fairly quickly into difficulties when attempting to account for non-trivial generalisations about the nature of human intellectual competencies.

The most significant of these is the problem of how different intelligences intertwine in real-life applications, for `nearly all cultural roles exploit more than one intelligence; at the same time, no performance can come about simply through the exercise of a single intelligence' (Gardner, 1993a, p. 208). An example here is tool manufacture and use, where both human and non-human creatures are known to be capable of making or deploying various objects to remedy problems (see Visalberghi, 1993, pp. 138-150). Such skills require not only visual-spatial, but also bodily-kinesthetic competence. More complex activities, such as devising new inventions, may require amalgamating these and other possible forms of capability, such as logico-mathematical skill, for example (see Gardner, 1993a, pp. 184-185; 233). Hence `one typically encounters complexes of intelligences functioning together smoothly, even seamlessly in order to execute intricate human activities' (Gardner, 1993a, p. 281). However Gardner is unable to explain satisfactorily how this occurs, for wherever such integrative complexities are raised, MI theory strains to elucidate how they take place, particularly in instances of `higher-level' cognitive operations like commonsense, originality, and metaphorical capacity, etc. (see Gardner, 1993a, pp. 288-295, 363). The overall result is a rather superficial and descriptive account, where an explication of the complexities raised by the theory dwindles at just the point where a richer, more substantive, and realistic scientific account of intelligent cognition goes begging. Consequently the theory's epistemic resources prove to be too meagre to the task of furnishing the requisite explanatory machinery for adequately sustaining its most general claims.

Difficulties are also to be found within a particular frame of intelligence. For example, the development of linguistic competence requires the interweaving of different components, some of which are not exclusive to the linguistic domain alone. Thus the syntactic and phonological cores of linguistic intelligence appear to be special and specific to the biological constitution of human beings, whereas other semantic and pragmatic aspects of language may exploit more general human information-processing mechanisms that are less exclusively tied to language modules in the brain. Hence it might be said that `syntax and phonology lie close to the core of linguistic intelligence while semantics and pragmatics include inputs from other intelligences', such as logical-mathematical and personal intelligences (Gardner, 1993a, p. 80). Consequently Gardner is forced to hedge on the question as to whether linguistic intelligence really is an intelligence apart, and concedes instead that language emerges only as a `relatively autonomous intelligence' (Gardner, 1993a, p. 89). Thus, aside from labouring to explain how the intelligences work together, MI theory also shows a disintegrative propensity, for the autonomy of the various intelligences now appear to be at risk of collapsing.

These methodological impediments therefore raise questions about the integrity of the theory's most basic categories, namely the intelligences themselves, since theory-laden observational evidence of functional specialisation in the brain constitutes only one of many possible constraints on theory construction. In particular, it may be asked whether the move from the identification of a set of modular core competencies in the brain, to the positing of a set of discrete and autonomous intellectual capacities, is methodologically sound, or whether, in the light of Gardner's admission to performing a `subjective' factor analysis, it simply begs the question in favour of some a priori framework. Furthermore, in recognising intertwining intelligences and interweaving intellectual components, Gardner may be giving credence to knowledge that is more reliable than the empiricist theory of knowledge underwriting MI theory would itself admit as being valid. Consequently legitimate doubts may be raised as to whether the classificatory framework that Gardner constructs actually hews the intellectual nature of creatures like us at our cognitive junctions.

Such concerns arise particularly in relation to the larger set of competencies that stem from interactions, for `in the normal course of events, the intelligences actually interact with, and build upon, one another from the beginning of life' (Gardner, 1993a, p. 280), and `once one begins to consider combinations of intelligences, one encounters an even larger set of ways in which an individual can be competent' (Gardner, 1993a, p. 320). Thus, although autonomous, the intelligences are not mutually exclusive (Gardner, 1993a, p. 365), for some intellectual competencies may generalise and interact with other domains (see Gardner, 1993a, p. 202). This attribute is noted when Gardner describes the complex union of skills that are manifested when, for example, Eskimos negotiate their way around open spaces, or Puluwat sailors navigate between islands (see Gardner, 1993a, pp. 184, 202-204, 338-341). However, if competence broadens out to encompass almost any form of acquired ability, however formed, then the question arises as to why the posited intelligences should necessarily be restricted to narrow and functionally specialised information-processing modules in the brain. The answer, of course, is that empiricist restrictions on evidence curtail acceptable claims to those that, in principle, can be verified directly or indirectly by observational means. However such methodological constraints on theory content and structure are overly severe, and a relaxation of admissible evidence along naturalistic and coherentist lines would permit a more generous and richer account of cognitive competence to be developed and sustained, which would accommodate a much wider and more realistic range of intellectual characteristics. Furthermore, since all knowledge is conjectural, provisional, and theoretical in nature, any framework positing a more or less stable taxonomy of `natural kinds' of intellectual capabilities or intelligences must remain open to being revised or superseded by possibly quite different classes of such kinds, as scientific knowledge of our mental capacities grows over time. Thus our access to natural kinds is not rigid, or set in stone, but is instead fluid, uncertain, and problematic (see Churchland, P.M., 1989, pp. 281-295).

Another difficulty can be seen in Gardner's treatment of the personal intelligences which, as a result of methodologically hiving off facts from values, are divided into two subtypes, namely interpersonal and intrapersonal forms of knowledge. Intrapersonal intelligence concerns the development of the internal aspects of a person, in which the core capacity is access to one's own feelings, affects, and emotions, and the capacity to discriminate among them. The origin of this `emotion-centred' intelligence can be discerned in the directly experienced feelings of the individual. Interpersonal intelligence, on the other hand, concerns the external aspects of the personal intelligences, in that it turns outwards towards other individuals, and its core capacity is the ability to notice and make distinctions among other individuals, particularly with regard to reading moods, temperaments, motivations, and intentions, etc. The origin of this `surface' form of intelligence lies in the direct perception of significant others. One's overall sense of self, therefore, is an amalgam, or combinatorial fusion, of the two forms of personal intelligences (see Gardner, 1993a, pp. 228, 240-244, 276, 1993b, pp. 22-26).

Since the patterns of development, end-states, and breakdown are far more varied in the personal intelligences than in the other forms of intellectual competence, they constitute a more attenuated and less compelling category in the overall MI framework, and may not prove to be so completely cognate with the other types of intelligences that Gardner identifies. Both forms of intelligence are also less independent of one another, and are more closely linked conceptually. Hence Gardner concedes that there is some ambiguity and uncertainty about the status of the personal intelligences in relation to the other intelligences, and wonders whether an individual's sense of self may constitute a higher and more integrated form of knowledge that acts as a `second-level regulator', or as an emergent capacity which modulates and exploits an individual's other intellectual capacities (see Gardner, 1993a, pp. 241, 268, 276, 296-298). However, such qualifiers raise questions about the overall coherence of MI theory, since anomalies concerning the status of some intelligences vis-a-vis other identified types, suggests that there are inconsistencies in how the criteria for identifying an intelligence are applied. A related question, and one that touches on a central epistemic virtue of theorising, is the matter of comprehensiveness, in which preference is to be accorded to those theories whose resources are able to explain more phenomena rather than less, with fewer anomalies, counter examples, and falsifying instances than the converse. Correspondingly, economy or simplicity is to be preferred to complexity, in that the least amount of explanatory apparatus is used to account for the widest range of phenomena. In addition, theories are to be preferred when they do not outrun their own explanatory resources by positing distinctions--in this case entertaining possible higher orders of intelligence---for which there is, on their own terms, no evidence (see Evers & Lakomski, 1991, p. 9).

Further methodological difficulties arise in connection with the symbol systems approach, which is used to furnish the various intelligences with cultural content. According to Gardner, the intelligences can only be implemented in symbol systems of meaning, which provide a scheme of interpretation, and give the intelligences their characteristic form (Gardner, 1993a, pp. 275, 299). However, although Gardner places considerable emphasis on the important role that symbols play in intelligent cognition, he nevertheless provides no coherent or conceptually unified account of what a symbol is, other than asserting that symbol systems are `culturally contrived systems of meaning which capture important forms of information' (Gardner, 1993a, p. 66). Nevertheless, when discussing intellectual assessment, Gardner suggests tacitly that what individuals learn from life is the ability to recognise a diverse range of patterns. This implicit formulation of what a symbol consists of accords well with the Physical Symbol Systems Hypothesis that Vera and Simon articulate, for instance, in relation to the situated cognition approach to artificial intelligence, in which symbols are presumed to be `patterns of neuronal arrangements of some kind' (see Vera & Simon, 1993, pp. 8-10). However, although symbolic structures may include language, pictures, diagrams, numbers, rituals, gestures, stories, and other `symbolic vehicles of thought', representation, and communication (see Gardner, 1993a, pp. 25, 66, 302-303, 1993b, p. 16; Vera & Simon, 1993, p. 10), exactly what pattern processing consists of is not specified. Consequently, when discussing the nature of bodily-kinesthetic intelligence, for example, Gardner talks of `kinesthetic thought' as being symbolised in a `kinesthetic language'. But, what that kinesthetic language may be, or how it may be represented or computed in the brain's substrates, is not explained, except that Gardner suggests vaguely that it is some experience or memory of how our bodies might feel in performing a sensorimotor task in some particular context or other (see Gardner, 1993a, p. 229).

In the case of spatial intelligence, representation and computation of information appear to consist of a number of loosely related capacities that operate as a family, such as the ability to recognise instances of similar elements, to transform one element into another, to produce a graphic likeness of spatial information that grows out of one's observations of the visual world, and so on. According to Gardner, `both introspective and experimental evidence suggest that the preferred mode for solution of "imagery problems" is through the positing of an internal mental image which can then be manipulated in ways that parallel the operations of the workaday world' (Gardner, 1993a, p. 175). However there is no privileged relationship between visual input and spatial intelligence, since blind individuals may also develop spatial capacities, and therefore `spatial representational systems are equally accessible to visual or to tactile experience' (Gardner, 1993a, pp. 174, 185-186). Thus representation in the spatial domain is heterogeneous in nature. However, in the absence of a cohesive explication of what a symbol is in this context, claims of susceptibility to encoding in a symbol system being integral to spatial and other forms of cognition, and of human intelligence possibly having `natural' gravitation towards embodiment in a symbolic system, are statements that appear to be vague, if not vacuous, in meaning (see Gardner, 1993a, p. 66). Furthermore resort to introspective sources of evidence for internal brain states is methodologically suspect, for our reflections into the nature of our own mental states are just as theory-laden as our perceptions of the external world, and are known to be as unreliable (Churchland, P.M., 1989, p. 140). Since we have no privileged access to our internal awareness (see Nisbett & Wilson, 1977, pp. 231-259; Wilson, Hull, & Johnson, 1981, pp. 53-71), and short of defensible theories of how our cognitive kinematics and dynamics are organised and take place, we at best confabulate about the nature of our subjectivity (Churchland, P.S., 1983, p. 83).

The shortcomings of Gardner's symbol-systems approach to cognition are particularly apparent in his efforts to account for the problem of how `higher level' cross-modal connections between the various intelligences are realised, which make use of seemingly broad and general mental skills that cut across the individual intelligences in a synergistic or interactive manner. One manifestation of this problem occurs when MI theory attempts to account for the development of symbolic capacities, since `certain aspects of development do not fall easily into the "pure intelligence" mode of analysis' (Gardner, 1993a, p. 314). In particular, MI theory is unable to explain why `waves' of symbolic and intellectual development are not confined to their own particular symbolic streams of intellectual capacity, in that they wash over the demarcating boundaries between the intelligences (see Gardner, 1993a, pp. 307-311). Given the paucity of adequate theoretical resources for explaining how the intelligences relate to one another, it is perhaps unsurprising to find that, despite a lack of evidence, Gardner is prepared to entertain the possibility of `amodal' or `cross-modal' capacities, such as metaphoric ability for example, as separate forms of intelligence in their own right (see Gardner, 1993a, pp. 294, 315). However, this temptation to introduce ad hoc or auxiliary hypotheses to cover inconvenient gaps in the theory's explanatory fabric is methodological bad air, for despite appearances, such hypotheses do not rescue other generalisations in the theory from their explanatory and predictive inadequacies, but instead only worsen the vices the framework already has, by highlighting its violations of conservatism, simplicity, generality, and conceptual unity, as key epistemic criteria essential to coherent theory construction (see Churchland, P.M., 1989, pp. 179-181; Quine & Ullian, 1970, pp. 49-50).

Missing cognitive matter

Gardner's account of the multiple intelligences follows the general methodological contours of the `classical' or `good old-fashioned' (GOFAI) approach to artificial intelligence research in traditional cognitive science (see Churchland, P.M., 1989, pp. xiii, 171-172; Coveney & Highfield, 1996, pp. 127-130). Central to this research program is the Physical Symbol Systems Hypothesis, which is an information-processing approach developed in the work of Allen Newell and Herbert Simon (see Dunlop & Fetzer, 1993, p. 98; Evers & Lakomski, 2000, pp. 10-11). The central idea is that intelligent cognitive processes consist of computations in the brain, or in a suitably programmed computer, which operate on symbolic representations or structures that have syntactic properties of some kind or other. According to this hypothesis, the necessary and sufficient conditions for something to be capable of intelligent action, or to have mental properties, or even a mind for that matter, is that such a system be a physical symbol system. In accordance with this conception, a suitably configured computer, such as the Turing machine, was thought to possess the necessary and sufficient conditions to have an intelligent mind, and the so-called Turing test was designed, and has long been used, in efforts to demonstrate the validity of this hypothesis.

The distinguishing property of mind, therefore, resides in the software program that runs in the physical computational system in question, whether these are composed of neurones or silicon circuits. Because of very weak constraints placed on theories of mind by logical empiricism, the physical substance used to encode and implement symbolic processing was considered unimportant. All that mattered is that the system doing the computing be functionally adequate to the tasks of transforming observed inputs into required outputs. On this view, the task of understanding intelligent cognition and action amounted to placing an emphasis not on the psychological properties of the system being studied, but on the formal properties of the symbolic representations themselves (see Evers & Lakomski, 2000, p. 11). The methodological upshot of this point of view has been the so-called `top-down' approach to cognitive science which, in its relative independence and isolation from questions of physical implementation, has largely disconnected itself from constraints that scientific knowledge relating to brain structure and function might impose on theorising and research (see Churchland, 1989, p. 156).

In fairness, Gardner considers neuroscientific discoveries that have illuminated the functional architecture of the brain, the so-called information-processing modules that constitute the biological cores of his intelligences. If the computer analogy is invoked, which divides the mind/brain into three levels of organisation, these findings relate to the lowest level of structural implementation. At the highest semantic level of organisation are matters of cultural significance, and their associated symbolic representations. However it is at the middle syntactic level where Gardner's account of the multiple intelligences falls fiat, for he is unable to explicate coherently how the algorithms, which carry out intelligent computations, are realised. It is at this level where substantive theorising is absent, and cognitive matter is missing from Gardner's overall conception. This is significant, because it has important educational ramifications. Without a coherent and realistic theory of mind and cognition, capable of explicating processes of perception, recognition, memory, learning, reasoning, reflection, introspection, subjectivity, selfhood, and intelligent adaptation, and so on, Gardner's hopes of informing better educational practices, and his claim to offer an explanation for educational success and failure (see Gardner, 1993a, p. 333), will amount to little, and the theory of multiple intelligences will go the way of other similar contenders in the history of science, having come to grief on the rocks of explanatory inadequacy.

Nevertheless, from the beginning, research efforts based on the computer analogy, along with the classical AI research program from whence it arises, have been continuously thwarted and frustrated by many obstacles and intractable empirical problems. These include failure to emulate realistically the cognitive and behavioural skills of humans, and other non-linguistic animals, in effortlessly recognising and responding to patterns embedded in complex and noisy stimulus fields; and the brittleness and inflexibility that classical AI systems manifest in coping satisfactorily with imperfect, partial, or ambiguous information. Furthermore classical AI systems have been unable to accommodate the subtlety and complexity of context-dependent knowledge, which effectively limits them to very restricted and narrow domains of application (see Churchland, P.M., 1989, pp. 155-156; Coveney & Highfield, 1996, pp. 126-130).

As to the question of how many levels of organisation there are in the brain, the answer is not known, but it would appear that there are many more than the three that are posited by the classical approach to AI. However this is a matter to be settled through empirical inquiry, and not via a priori theorising (see Churchland, P.S., 1992, p. 359). Nevertheless clues to how intelligent cognition is organised in the substrates of the brain may be found in a new bottom-up approach to cognitive science known as connectionist AI, which as an alternative research framework already shows an impressive degree of fertility and utility, which has hitherto not been realised by the classical approach to AI (see Churchland, P.M., 1989, 1995; Churchland, P.S. & Sejnowski, 1994; Coveney & Highfield, 1996).

The new cognitive science

Over the past couple of decades, a class of computer programs known as artificial neural networks (ANNs) have been developed and used in artificial intelligence research for investigating the structural and functional properties of the brain. As analogues of biological networks, these heuristic devices link the microscopic level of neurobiology with higher systemic levels of organised cognition and behaviour, and permit investigators to examine how intelligent properties emerge from the organised interaction of discrete lower level neuronal components, which do not have such properties. In this way, connectionist AI research makes coherent contact with neurobiological constraints, and illuminates a conception of cognitive kinematics and dynamics that encompasses not only features of the human brain, but also the brains of creatures throughout the animal world. The computational explanations sought refer to the informational content of the physical signals found in the brain, and research aims to uncover the algorithms that are implemented in, and constrained by, its architecture, thereby enabling analysis of the underlying principles of brain function (see Sejnowski, Koch, & Churchland, P.S., 1988, pp. 1299-1301; Smolensky, 1988, pp. 9-10).

ANNs are composed of a number of layers of neuron-like units that carry out simultaneous parallel distributed processing (PDP) tasks as incoming stimuli-vectors are propagated through the network, and are transformed as they converge on a calibrated or trained output vector. Each vector, as it moves from input layer through to a hidden middle level layer (or series of such layers), and on through to an output layer in the network, is instantiated by an activation profile of connection weight values, which are a function of specific contextual features or dimensions in the overall input and its particular transforms, as the vector moves through the network to a particular learned output (see Churchland, P.M., 1989, pp. 160-163). The representation that the network has of the input stimulus, and which it transforms into some other cognitive or behavioural output vector as it is processed in either a feedforward or recurrent manner, consists of the temporal combination of the various levels of activation across all of the individual input units of the network. Thus the network's knowledge is represented in the configuration of connection weight values that the network has calibrated during learning. PDP therefore permits a powerful combinatorial system of representational coding, which allows the representing network to make very fine-grained sensory and cognitive discriminations that are highly contextual, dynamically continuous, and subsymbolic in nature (see Churchland, P.M., 1989, pp. 163-164, 1995, pp. 21, 97-121; Smolensky, 1988, pp. 1-74), and which as patterns of activation can be rendered in an analytically tractable form.

Analysis of the various weight configurations in the hidden or middle layer(s) of a trained ANN reveals that the system is capable of extracting and capturing higher order statistical relationships from the details of the input stimulus, and of partitioning activation vector spaces into fecund and highly structured category hierarchies, which are organised around a similarity metric. The highest partition corresponds to the simplest and most global feature, or set of features, that both unite and divide the variety of other subordinate categories found at lower levels of organisation, and which find their ultimate representation in the chaotic dimensions of the input itself. Thus ANNs are capable of developing a system of internal representations that correspond to important features in the input they receive from their external environments, which are not explicitly represented in the inputs themselves. Vectors that cluster around a particular category sub-volume of weight space constitute prototypical instances of the relevant features found in the stimulus input (see Churchland, P.M., 1989, pp. 168, 172). Hence activation vectors having some complex of relational or structural features in common will cluster together in the same region of weight space to form a prototypical hot spot, which makes the network extremely sensitive to similarities along all relevant stimulus dimensions (Churchland, P.M., 1989, pp. 204-205). In becoming `tuned' to the weave of features making up the input stimuli, the network is then able to use the `knowledge' it has acquired and stored in its connections to generalise its responses successfully to new cases (see Churchland, P.M., 1989, p. 167). The process of category structuring along a similarity metric, and of appropriately responding to prototypical instances of such categories, are functions that are typical of theories (Churchland, P.M., 1989, p. 177), for the whole point of partitioning the hidden unit layer(s) into activation vectors, or a conceptual framework of some sort, `is to produce and steer well-tuned behaviour' (see Churchland, P.M., 1995, pp. 90-91). Hence discoveries in connectionist Al research suggest a conception of cognition and behaviour in which the principal form of representation is a high-dimensional activation vector, and the principal form of computation is vector-to-vector transformation, effected by a matrix of differently weighted synapses (see Churchland, P.M., 1989, pp. 160-164, 209, 1995, p. 508). From this perspective, the term intelligence `is itself a vector' (Churchland, P.M., 1995, p. 253), the specific content and semantics of which can be characterised by geometric and mathematical neuromodeling (see Churchland, P.M., 1998, pp. 5-32).

As such, ANNs show themselves to be capable of learning, representing, and performing a wide range of sophisticated tasks which accurately imitate computational functions that various parts of the intelligent brain might be expected to perform. These include, for instance: Fusion.net, which imitates properties of human stereoscopic vision; an acoustic network that is capable of performing sonar analysis and correctly discriminating explosive-mine and harmless-rock echoes; NETtalk, which is a network that is capable of producing phonetic outputs in response to read inputs; a crab network that is capable of executing coordinated sensorimotor routines; and EMPATH, which is a network that is capable of reading emotional states from photographs of human faces (see Churchland, P.M., 1995, pp. 71-96, 125-128). All of these networks mimic the kind of complex computations that the various computational cores of the multiple intelligences might be expected to perform, and the algorithms by which these feats of representation and computation are achieved can be explicated via the theoretical machinery afforded by computational neuroscience. Consequently the framework and the tools of the new cognitive science suggest a more fruitful avenue for exploring and understanding the intelligent capacities of the brain, human and nonhuman alike.

Conclusion

Analysis of Gardner's theory of multiple intelligences uncovers a conception that, in terms of its content and structure, is informed by an outmoded and narrow empiricist theory of knowledge, and associated theory of mind and cognition. As a consequence, the theory fails to furnish methodological resources that are adequate to the tasks of explicating and defending the integrity of its most basic categories, namely the multiple intelligences themselves, and of underwriting the explanatory generalisations that arise therefrom.

The suggested alternative, which overcomes the limitations of empiricism, is to treat matters of epistemology and methodology naturalistically, by looking to the best and most advanced current developments in cognitive scientific theorising and research for insights into the nature of intelligent cognition. Such an approach promises to provide much richer and more substantial resources for building realistic theories of intelligence, which more adequately informs the details of educational practice, as well as other domains of social life.

Keywords
cognitive processes        creativity      intelligence
computers developmental    psychology      knowledge representation

References

Ayer, A.J. (1975). Language, truth and logic. Harmondsworth:Penguin Books.

BonJour, L. (1985). The structure of empirical knowledge. Cambridge, MA: Harvard University Press.

Churchland, P.M. (1989). A neurocomputational perspective: The nature of mind and the structure of science (A Bradford Book). Cambridge, MA: MIT Press.

Churchland, P.M. (1995). The engine of reason, the seat of the soul: A philosophical journey into the brain (A Bradford Book). Cambridge, MA: MIT Press.

Churchland, P.M. (1998). Conceptual similarity across sensory and neural diversity: The Fodor/Lepore challenge answered. Journal of Philosophy, 95 (1), 5-32.

Churchland, P.S. (1983). Consciousness: The transmutation of a concept. Pacific Philosophical Quarterly, 64, 80-95.

Churchland, P.S. (1987). Epistemology in the age of neuroscience. Journal of Philosophy, 84 (10), 544-553.

Churchland, P.S. (1992). Neurophilosophy: Toward a unified science of the mind/brain (A Bradford Book). Cambridge, MA: MIT Press.

Churchland, P.S. & Sejnowski, T.J. (1994). The computational brain (A Bradford Book). Cambridge, MA: MIT Press.

Coveney, P. & Highfield, IL. (1996). Frontiers of complexity: The search for order in a chaotic world. London: Faber & Faber.

Dunlop, C.E.M. & Fetzer, J.H. (1993). Glossary of cognitive science. New York: Paragon House.

Evers, C.W. (1988). Educational administration and the new philosophy of science. Journal of Educational Administration, 26 (1), 3-22.

Evers, C.W. (1995). Recent developments in educational administration. Leading and Managing, 1 (1), 1-13.

Evers, C.W. & Lakomski, G. (1991). Knowing educational administration: Contemporary methodological controversies in educational administration research. Oxford: Pergamon Press.

Evers, C.W. & Lakomski, G. (1993). Justifying educational administration. Educational Management and Administration, 21 (3), 140-152.

Evers, C.W. & Lakomski, G. (1996). Science in educational administration: A post-positivist conception. Educational Administration Quarterly, 32 (3), 379-402.

Evers, C.W. & Lakomski, G. (2000). Doing educational administration: A theory of administrative practice. Oxford: Pergamon Press.

Gardner, H. (1993a). Frames of mind: The theory of multiple intelligences (2nd ed.). New York: Basic Books.

Gardner, H. (1993b). The multiple intelligences: The theory in practice. New York: Basic Books.

Gardner, H. (1994). Multiple intelligences theory. In R.J. Sternberg (Ed.), Encyclopedia of human intelligence (Vol. 2). New York: Macmillan.

Gardner, H. (1998). A multiplicity of intelligences. Scientific American, 9 (4), 18-23.

Gottfredson, L.S. (1998). The general intelligence factor. Scientific American, 9 (4), 24-29.

Matthews, D. (1995) Gardner's multiple intelligence theory: An evaluation of relevant research literature and a consideration of its application to gifted education. In R. Fogarty & J. Bellanca (Eds.), Multiple intelligences: A collection. Highett, Vic.: Hawker Brownlow Education, Australia.

Nisbett, R.E. & Wilson, T.D. (1977). Telling more than we can know: Verbal reports on mental processes. Psychological Review, 84 (3), 231-259.

Quine, W.V.O. (1951). Two dogmas of empiricism. Philosophical Review, 60, 20-43.

Quine, W.V.O. (1969). Epistemology naturalized. In Ontological relativity and other essays. New York: Columbia University Press.

Quine, W.V.O. & Ullian, J.S. (1970). The web of belief New York: Random House.

Sejnowski, T.J., Koch, C., & Churchland, P.S. (1988). Computational neuroscience. Science, 241, 1299-1306.

Smolensky, P. (1988). On the proper treatment of connectionism. Behavioral and Brain Sciences, 11, 1-74.

Vera, A.H. & Simon, H.A. (1993). Situated action: A symbolic interpretation. Cognitive Science, 17, 7-48.

Visalberghi, E. (1993). Capuchin monkeys: A window into tool use in apes and humans. In K.R. Gibson & T. Ingold (Eds.), Tools, language and cognition in human evolution. Cambridge: Cambridge University Press.

Williams, M. (1977). Groundless belief: An essay on the possibility of epistemology. Oxford: Basil Blackwell.

Williams, M. (1980). Coherence, justification, and truth. Review of Metaphysics, 34, 243-272.

Wilson, T.D., Hull, J.G., & Johnson, J. (1981). Awareness and self-perception: Verbal reports on internal states. Journal of Personality and Social Psychology, 40 (1), 53-71.

Dr Nicholas M. Allix works in the Centre for Organisational Learning and Development, Faculty of Education, The University of Melbourne, Parkville, Victoria 3010.