PAGE 13

The main argument for representationalism appeals to the transparency of experience (cf. Tye 2000: 45-51). The properties that characterize what it's like to have a perceptual experience are presented in experience as properties of objects perceived: in attending to an experience, one seems to ‘see through it’ to the objects and properties it is experiences of. They are not presented as properties of the experience itself. If nonetheless they were properties of the experience, perception would be massively deceptive. But perception is not massively deceptive. According to the representationalist, the phenomenal character of an experience is due to its representing objective, non-experiential properties. (In veridical perception, these properties are locally instantiated; in illusion and hallucination, they are not.) On this view, introspection is indirect perception: one comes to know what phenomenal features one's experience has by coming to know what objective features it represents.


In order to account for the intuitive differences between conceptual and sensory representations, representationalists appeal to their structural or functional differences. Dretske (1995), for example, distinguishes experiences and thoughts on the basis of the origin and nature of their functions: an experience of a property 'P' is a state of a system whose evolved function is to indicate the presence of 'P' in the environment; a thought representing the property 'P', on the other hand, is a state of a system whose assigned (learned) function is to calibrate the output of the experiential system. Rey (1991) takes both thoughts and experiences to be relations to sentences in the language of thought, and distinguishes them on the basis of (the functional roles of) such sentences' constituent predicates. Lycan (1987, 1996) distinguishes them in terms of their functional-computational profiles. Tye (2000) distinguishes them in terms of their functional roles and the intrinsic structure of their vehicles: thoughts are representations in a language-like medium, whereas experiences are image-like representations consisting of ‘symbol-filled arrays.’ (the account of mental images in Tye 1991.)

Phenomenalists tend to make use of the same sorts of features (function, intrinsic structure) in explaining some of the intuitive differences between thoughts and experiences; but they do not suppose that such features exhaust the differences between phenomenal and non-phenomenal representations. For the phenomenalism, it is the phenomenal properties of experiences ~ qualia themselves ~ that constitute the fundamental difference between experience and thought. Peacocke (1992), for example, develops the notion of a perceptual ‘scenario’ (an assignment of phenomenal properties to coordinates of a three-dimensional egocentric space), whose content is ‘correct’ (a semantic property) if in the corresponding ‘scene’ (the portion of the external world represented by the scenario) properties are distributed as their phenomenal analogues are in the scenario.

Another sort of representation championed by phenomenalists (e.g., Block, Chalmers (2003) and Loar (1996)) is the ‘phenomenal concept’ ~ a conceptual/phenomenal hybrid consisting of a phenomenological ‘sample’ (an image or an occurrent sensation) integrated with (or functioning as) a conceptual component. Phenomenal concepts are postulated to account for the apparent fact (among others) that, as McGinn (1991) puts it, ‘you cannot form [introspective] concepts of conscious properties unless you yourself instantiate those properties.’ One cannot have a phenomenal concept of a phenomenal property 'P', and, hence, phenomenal beliefs about P, without having experience of 'P', because 'P' itself is (in some way) constitutive of the concept of 'P'. (Jackson 1982, 1986 and Nagel 1974.)

Though imagery has played an important role in the history of philosophy of mind, the important contemporary literature on it is primarily psychological. In a series of psychological experiments done in the 1970s (summarized in Kosslyn 1980 and Shepard and Cooper 1982), subjects' response time in tasks involving mental manipulation and examination of presented figures was found to vary in proportion to the spatial properties (size, orientation, etc.) of the figures presented. The question of how these experimental results are to be explained has kindled a lively debate on the nature of imagery and imagination.

Kosslyn (1980) claims that the results suggest that the tasks were accomplished via the examination and manipulation of mental representations that themselves have spatial properties ~ i.e., pictorial representations, or images. Others, principally Pylyshyn (1979, 1981, 2003), argue that the empirical facts can be explained in terms exclusively of discursive, or propositional representations and cognitive processes defined over them. (Pylyshyn takes such representations to be sentences in a language of thought.)

The idea that pictorial representations are literally pictures in the head is not taken seriously by proponents of the pictorial view of imagery The claim is, rather, that mental images represent in a way that is relevantly like the way pictures represent. (Attention has been focussed on visual imagery ~ hence the designation ‘pictorial’; though of course there may imagery in other modalities ~ auditory, olfactory, etc. ~ as well.)

The distinction between pictorial and discursive representation can be characterized in terms of the distinction between analog and digital representation (Goodman 1976). This distinction has itself been variously understood (Fodor & Pylyshyn 1981, Goodman 1976, Haugeland 1981, Lewis 1971, McGinn 1989), though a widely accepted construal is that analog representation is continuous (i.e., in virtue of continuously variable properties of the representation), while digital representation is discrete (i.e., in virtue of properties a representation either has or doesn't have) (Dretske 1981). (An analog/digital distinction may also be made with respect to cognitive processes. (Block 1983.)) On this understanding of the analog/digital distinction, imagistic representations, which represent in virtue of properties that may vary continuously (such as being more or less bright, loud, vivid, etc.), would be analog, while conceptual representations, whose properties do not vary continuously (a thought cannot be more or less about Elvis: either it is or it is not) would be digital.

It might be supposed that the pictorial/discursive distinction is best made in terms of the phenomenal/nonphenomenal distinction, but it is not obvious that this is the case. For one thing, there may be nonphenomenal properties of representations that vary continuously. Moreover, there are ways of understanding pictorial representation that presuppose neither phenomenality nor analogicity. According to Kosslyn (1980, 1982, 1983), a mental representation is ‘quasi-pictorial’ when every part of the representation corresponds to a part of the object represented, and relative distances between parts of the object represented are preserved among the parts of the representation. But distances between parts of a representation can be defined functionally rather than spatially ~ for example, in terms of the number of discrete computational steps required to combine stored information about them. (Rey 1981.)

Tye (1991) proposes a view of images on which they are hybrid representations, consisting both of pictorial and discursive elements. On Tye's account, images are ‘(labelled) interpreted symbol-filled arrays.’ The symbols represent discursively, while their arrangement in arrays has representational significance (the location of each ‘cell’ in the array represents a specific viewer-centred 2-D location on the surface of the imagined object)

The contents of mental representations are typically taken to be abstract objects (properties, relations, propositions, sets, etc.). A pressing question, especially for the naturalist, is how mental representations come to have their contents. Here the issue is not how to naturalize content (abstract objects can't be naturalized), but, rather, how to provide a naturalistic account of the content-determining relations between mental representations and the abstract objects they express. There are two basic types of contemporary naturalistic theories of content-determination, causal-informational and functional.

Causal-informational theories (Dretske 1981, 1988, 1995) hold that the content of a mental representation is grounded in the information it carries about what does (Devitt 1996) or would (Fodor 1987, 1990) cause it to occur. There is, however, widespread agreement that causal-informational relations are not sufficient to determine the content of mental representations. Such relations are common, but representation is not. Tree trunks, smoke, thermostats and ringing telephones carry information about what they are causally related to, but they do not represent (in the relevant sense) what they carry information about. Further, a representation can be caused by something it does not represent, and can represent something that has not caused it.

The main attempts to specify what makes a causal-informational state a mental representation are Asymmetric Dependency Theories (e.g., Fodor 1987, 1990, 1994) and Teleological Theories (Fodor 1990, Millikan 1984, Papineau 1987, Dretske 1988, 1995). The Asymmetric Dependency Theory distinguishes merely informational relations from representational relations on the basis of their higher-order relations to each other: informational relations depend upon representational relations, but not vice-versa. For example, if tokens of a mental state type are reliably caused by horses, cows-on-dark-nights, zebras-in-the-mist and Great Danes, then they carry information about horses, etc. If, however, such tokens are caused by cows-on-dark-nights, etc. because they were caused by horses, but not vice versa, then they represent horses.

According to Teleological Theories, representational relations are those a representation-producing mechanism has the selected (by evolution or learning) function of establishing. For example, zebra-caused horse-representations do not mean zebra, because the mechanism by which such tokens are produced has the selected function of indicating horses, not zebras. The horse-representation-producing mechanism that responds to zebras is malfunctioning.

Functional theories (Block 1986, Harman 1973), hold that the content of a mental representation is grounded in its (causal computational, inferential) relations to other mental representations. They differ on whether relata should include all other mental representations or only some of them, and on whether to include external states of affairs. The view that the content of a mental representation is determined by its inferential/computational relations with all other representations is holism; the view it is determined by relations to only some other mental states is localism (or molecularism). (The view that the content of a mental state depends on none of its relations to other mental states is atomism.) Functional theories that recognize no content-determining external relata have been called solipsistic (Harman 1987). Some theorists posit distinct roles for internal and external connections, the former determining semantic properties analogous to sense, the latter determining semantic properties analogous to reference (McGinn 1982, Sterelny 1989)

(Reductive) representationalists (Dretske, Lycan, Tye) usually take one or another of these theories to provide an explanation of the (Non-conceptual) content of experiential states. They thus tend to be Externalists about phenomenological as well as conceptual content. Phenomenalists and non-deductive representationalists (Block, Chalmers, Loar, Peacocke, Siewert), on the other hand, take it that the representational content of such states is (at least in part) determined by their intrinsic phenomenal properties. Further, those who advocate a phenomenology-based approach to conceptual content (Horgan and Tiensen, Loar, Pitt, Searle, Siewert) also seem to be committed to internalist individuation of the content (if not the reference) of such states.

Generally, those who, like informational theorists, think relations to one's (natural or social) environment are (at least partially) determinative of the content of mental representations are Externalists (e.g., Burge 1979, 1986, McGinn 1977, Putnam 1975), whereas those who, like some proponents of functional theories, think representational content is determined by an individual's intrinsic properties alone, are internalists (or individualists; cf. Putnam 1975, Fodor 1981)

This issue is widely taken to be of central importance, since psychological explanation, whether commonsense or scientific, is supposed to be both causal and content-based. (Beliefs and desires cause the behaviours they do because they have the contents they do. For example, the desire that one have a beer and the beliefs that there is beer in the refrigerator and that the refrigerator is in the kitchen may explain one's getting up and going to the kitchen.) If, however, a mental representation's having a particular content is due to factors extrinsic to it, it is unclear how its having that content could determine its causal powers, which, arguably, must be intrinsic. Some who accept the standard arguments for externalism have argued that internal factors determine a component of the content of a mental representation. They say that mental representations have both ‘narrow’ content (determined by intrinsic factors) and ‘wide’ or ‘broad’ content (determined by narrow content plus extrinsic factors). (This distinction may be applied to the sub-personal representations of cognitive science as well as to those of commonsense psychology.

Narrow content has been variously construed. Putnam (1975), Fodor (1982)), and Block (1986), for example, seem to understand it as something like de dicto content (i.e., Frégean sense, or perhaps character, à la Kaplan 1989). On this construal, narrow content is context-independent and directly expressible. Fodor (1987) and Block (1986), however, have also characterized narrow content as radically inexpressible. On this construal, narrow content is a kind of proto-content, or content-determinant, and can be specified only indirectly, via specifications of context/wide-content pairings. On both construal, narrow contents are characterized as functions from context to (wide) content. The narrow content of a representation is determined by properties intrinsic to it or its possessor such as its syntactic structure or its intra-mental computational or inferential role (or its phenomenology.

Burge (1986) has argued that causation-based worries about externalist individuation of psychological content, and the introduction of the narrow notion, are misguided. Fodor (1994, 1998) has more recently urged that a scientific psychology might not need narrow content in order to supply naturalistic (causal) explanations of human cognition and action, since the sorts of cases they were introduced to handle, viz., Twin-Earth cases and Frége cases, are either nomologically impossible or dismissible as exceptions to non-strict psychological laws.

The leading contemporary version of the Representational Theory of Mind, the Computational Theory of Mind (CTM), claims that the brain is a kind of computer and that mental processes are computations. According to the computational theory of mind, cognitive states are constituted by computational relations to mental representations of various kinds, and cognitive processes are sequences of such states. The computational theory of mind and the representational theory of mind, may by attempting to explain all psychological states and processes in terms of mental representation. In the course of constructing detailed empirical theories of human and animal cognition and developing models of cognitive processes implementable in artificial information processing systems, cognitive scientists have proposed a variety of types of mental representations. While some of these may be suited to be mental relata of commonsense psychological states, some ~ so-called ‘subpersonal’ or ‘sub-doxastic’ representations ~ are not. Though many philosophers believe that computational theory of mind can provide the best scientific explanations of cognition and behaviour, there is disagreement over whether such explanations will vindicate the commonsense psychological explanations of prescientific representational theory of mind.

According to Stich's (1983) Syntactic Theory of Mind, for example, computational theories of psychological states should concern themselves only with the formal properties of the objects those states are relations to. Commitment to the explanatory relevance of content, however, is for most cognitive scientists fundamental (Fodor 1981, Pylyshyn 1984, Von Eckardt 1993). That mental processes are computations, which computations are rule-governed sequences of semantically evaluable objects, and that the rules apply to the symbols in virtue of their content, are central tenets of mainstream cognitive science.

Explanations in cognitive science appeal to a many different kinds of mental representation, including, for example, the ‘mental models’ of Johnson-Laird 1983, the ‘retinal arrays,’ ‘primal sketches’ and ‘2½ -D sketches’ of Marr 1982, the ‘frames’ of Minsky 1974, the ‘sub-symbolic’ structures of Smolensky 1989, the ‘quasi-pictures Kosslyn 1980, and the ‘interpreted symbol-filled arrays’ of Tye 1991 ~ in addition to representations that may be appropriate to the explanation of commonsense psychological states. Computational explanations have been offered of, among other mental phenomena, belief (Fodor 1975, Field 1978), visual perception (Marr 1982, Osherson, et al. 1990), rationality (Newell and Simon 1972, Fodor 1975, Johnson-Laird and Wason 1977), language learning and  (Chomsky 1965, Pinker 1989), and musical comprehension (Lerdahl and Jackendoff 1983).

A fundamental disagreement among proponents of computational theory of mind concerns the realization of personal-level representations (e.g., thoughts) and processes (e.g., inferences) in the brain. The central debate here is between proponents of Classical Architectures and proponents of Conceptionist Architectures.

The classicists (e.g., Turing 1950, Fodor 1975, Fodor and Pylyshyn 1988, Marr 1982, Newell and Simon 1976) hold that mental representations are symbolic structures, which typically have semantically evaluable constituents, and that mental processes are rule-governed manipulations of them that are sensitive to their constituent structure. The connectionists (e.g., McCulloch & Pitts 1943, Rumelhart 1989, Rumelhart and McClelland 1986, Smolensky 1988) hold that mental representations are realized by patterns of activation in a network of simple processors (‘nodes’) and that mental processes consist of the spreading activation of such patterns. The nodes themselves are, typically, not taken to be semantically evaluable; nor do the patterns have semantically evaluable constituents. (Though there are versions of Connectionism ~ ‘localist’ versions ~ on which individual nodes are taken to have semantic properties (e.g., Ballard 1986, Ballard & Hayes 1984).) It is arguable, however, that localist theories are neither definitive nor representative of the Conceptionist program (Smolensky 1988, 1991, Chalmers 1993).

Classicists are motivated (in part) by properties thought seems to share with language. Fodor's Language of Thought Hypothesis (LOTH) (Fodor 1975, 1987), according to which the system of mental symbols constituting the neural basis of thought is structured like a language, provides a well-worked-out version of the classical approach as applied to commonsense psychology. According to the language of thought hypothesis, the potential infinity of complex representational mental states is generated from a finite stock of primitive representational states, in accordance with recursive formation rules. This combinatorial structure accounts for the properties of productivity and systematicity of the system of mental representations. As in the case of symbolic languages, including natural languages (though Fodor does not suppose either that the language of thought hypothesis explains only linguistic capacities or that only verbal creatures have this sort of cognitive architecture), these properties of thought are explained by appeal to the content of the representational units and their combinability into contentful complexes. That is, the semantics of both language and thought is compositional: the content of a complex representation is determined by the contents of its constituents and their structural configuration.

Connectionists are motivated mainly by a consideration of the architecture of the brain, which apparently consists of layered networks of interconnected neurons. They argue that this sort of architecture is unsuited to carrying out classical serial computations. For one thing, processing in the brain is typically massively parallel. In addition, the elements whose manipulation drives computation in Conceptionist networks (principally, the connections between nodes) are neither semantically compositional nor semantically evaluable, as they are on the classical approach. This contrast with classical computationalism is often characterized by saying that representation is, with respect to computation, distributed as opposed to local: representation is local if it is computationally basic; and distributed if it is not. (Another way of putting this is to say that for classicists mental representations are computationally atomic, whereas for connectionists they are not.)

Moreover, connectionists argue that information processing as it occurs in Conceptionist networks more closely resembles some features of actual human cognitive functioning. For example, whereas on the classical view learning involves something like hypothesis formation and testing (Fodor 1981), on the Conceptionist model it is a matter of evolving distribution of ‘weight’ (strength) on the connections between nodes, and typically does not involve the formulation of hypotheses regarding the identity conditions for the objects of knowledge. The Conceptionist network is ‘trained up’ by repeated exposure to the objects it is to learn to distinguish; and, though networks typically require many more exposures to the objects than do humans, this seems to model at least one feature of this type of human learning quite well.

Further, degradation in the performance of such networks in response to damage is gradual, not sudden as in the case of a classical information processor, and hence more accurately models the loss of human cognitive function as it typically occurs in response to brain damage. It is also sometimes claimed that Conceptionist systems show the kind of flexibility in response to novel situations typical of human cognition ~ situations in which classical systems are relatively ‘brittle’ or ‘fragile.’