Published in The Philosophy of Artificial Intelligence, edited by Margaret Boden, in the Oxford Readings in Philosophy Series (Oxford University Press, 1990), pp. 368 - 440.
The Connectionist Construction of Concepts
The character of computational modelling of cognition depends on an underlying theory of representation. Classical cognitive science has exploited the syntax/semantics theory of representation that derives from logic. But this has had the consequence that the kind of psychological explanation supported by classical cognitive science is conceptualist: psychological phenomena are modelled in terms of relations that hold between concepts, and between the sensors/effectors and concepts. This kind of explanation is inappropriate for the Proper Treatment of Connectionism (Smolensky 1988). Is there an alternative theory of representation that retains the advantages of classical theory, but which does not force psychological explanation into the conceptualist mould? I outline such an alternative by introducing an experience-based notion of nonconceptual content and by showing how a complex construction out of nonconceptual content can satisfy classical constraints on cognition. The psychologically fundamental structure of cognition is not the structure that holds between concepts, but, rather, the structure within concepts. The theory of the representational structure within concepts allows psychological phenomena to be explained as the progressive emergence of objectivity. This can be modelled computationally by means of the computational processes of a perspective-dependence-reducing transformer. This device may be thought of as a generalisation of a cognitive map, which includes the processes of map-formation and map use. It forms computational structures which take nonconceptual contents as inputs and yield nonconceptual contents as outputs, but do so in a way which makes the resulting capacity of the system less and less dependent on any particular perspectives, yielding satisfactory performance from any point of view.
1.1 Two Cognitive Science Frameworks for a Solution to the Problem of Embodied Cognition
Cognitive science theories are theories of how physical systems think. But a framework for cognitive science theorising must explain how it is possible for physical systems to think. How can intentional phenomena be part of the same world which is described by the natural sciences? How can there be organisms in the world which are capable of thinking about the world? How can the world include, as a part of itself, perspectives on the world? I shall call the problem of possibility introduced by these questions, “the problem of embodied cognition".
This article is about solutions to the problem of embodied cognition, which are both psychological and computational in character. The Language of Thought (LOT) framework (Fodor 1976, 1987, and see §3) is exhibited as a candidate solution, and a rival cognitive science framework (“C3" for Connectionist Construction of Concepts) developed. Both LOT and C3 serve also as methodologies for work in cognitive science, helping to direct research and to understand its significance. Hence two contrasts emerge from the paper: a contrast between two general conceptions of the enterprise of cognitive science and a contrast between two ways of understanding how embodied cognition is possible.
1.2 A Theory of Representation as a Means for Deriving Psychological Explanations from Computational Models.
A computational artefact which is held to have significance for psychological explanation is a “model". A model is just a physical object. How are psychological explanations to be extracted from it?
A cognitive science theory (“a theory") is a structured articulation of psychological explanations based on the functioning of the model. Cognitive science theorising thus rests on a conception of the relation between computational artefacts and psychological explanations. This relation is mediated by a theory of representation.
A representation is itself a physical object which has two kinds of properties: properties of the representational “vehicle" and properties of the representational “content". For example, a sequence of marks on a marking surface may be a representation. The alphanumeric letter sequencing that these marks instantiate is a property of the representational vehicle. And if the sequence happens to be the following, “Stanford is warmer than Oxford", then the content of the representation is that Stanford is warmer than Oxford. The representational vehicle is the medium that carries the representational content as its message.
In a model, the properties of a representational vehicle are all properties which have computational impact (for example, syntactic properties of LISP code). They are properties which affect the computational functioning of the model. And the properties which form the representational content are all properties which have psychological impact (for example, the task domain semantic properties of the LISP code). They are properties which affect the psychological explanations which can be derived from the model. So, on one side, the properties of a representation have a role in psychological explanation, and on the other side, they have a role in the computational functioning of the model. It is the theory of representation which must tie together these two sets of properties, and hence establish the connection between computational functioning and psychological explanation. It is a theory of representation which allows us to extract psychological import from computational physical objects; that gets a theory out of a model.
If cognitive science involves getting a psychological theory out of a computational model, and if a theory of representation is the way to do this, then in order to understand the nature of cognitive science theorising we need to understand the relation between computation, representational vehicles, representational content, and psychological explanation. The task is inherently multidisciplinary:
Figure 1: Four Levels of Analysis in Cognitive Science
1.3 Cognitive Science Frameworks
A cognitive science framework consists of an analysis at each of the four levels in figure 1. A central theme of this paper is that the analyses are not independent of each other. For example, given a Von Neumann analysis of the computational architecture of a model, and a syntactic analysis of the model's representational vehicles, the theory of content for the model would have to be a semantic theory. And the choice of semantic contents entails that a particular kind of psychological explanation (conceptualist explanation) is derived from the model (as will be explained later). Or, if one chooses a connectionist computational architecture, one may be lead, as for example Smolensky (1988) has been, to reject syntactic representational vehicles. As is shown in this article, this consequence should itself have implications for the kinds of contents which connectionist representational vehicles can carry, and thus implications for the kinds of psychological explanation which can be extracted from connectionist models. The analysis at each level constrains the analysis at the adjacent levels, so consequences can also be traced in a top-down direction.
A cognitive science framework, then, involves a decision at each of these levels, so that the decision at each level is compatible with the decisions at all of the other levels. Diagramming the possible choices at each level provides a representation of competing cognitive science frameworks. The terms which denote the choices in figure 2 are explained in the body of the article. But it may help to begin with the diagram:
Figure 2: LOT and C3 as Cognitive Science Frameworks
[Vertical arrows may be read as “constrains". Horizontal arrows indicate some possible kinds or choices at each level. The left hand side of the diagram is labelled as the set of choices which constitute the Language of Thought framework for cognitive science, and the right hand side is labelled as the set of choices which constitute the C3 framework for cognitive science. The unfamiliar terms in the diagram are explained in the body of the text.]
1.4 The Strategy of the Paper
An alternative kind of content from that presupposed by LOT is suggested, and the consequences of its use are considered for psychological explanation, for theories of representation, and for computational implementation.
If the alternative framework is to be a genuine alternative to LOT, then it must provide a solution to the problem of embodied cognition; it must indicate how the physical embodiment of cognition is possible. §(2) explains why this problem is a problem and provides a necessary and sufficient condition for its solution.
In §(3), I explain why the LOT interpretation of cognitive science offers a candidate solution to the problem of embodied cognition. I point out that this status depends on the computational use of the classical syntax/semantics theory of representation (S/S theory).
The dependence of LOT theorising on the S/S theory of representation entails that psychological modelling based on LOT employs conceptual content. So in §(4), I explain the distinction between conceptual content and nonconceptual content, give several examples which indicate the psychological need for a notion of nonconceptual content, and introduce a particular kind of nonconceptual content: construction-theoretic content (CTC).
In §(5), I explore the cognitive psychological consequences of modelling in terms of conceptual content. This establishes the contrast for §(6) to develop the idea that the nonconceptualist psychological task is to explain the cognitive emergence of objectivity. §(7) makes the notion of objectivity more precise, and provides a way of assessing any system for the degree to which it is a concept-exercising system. §(8) develops the connection between objectivity and perspective-independence.
§(9) shows how a psycho-computational theory of map-like transformations of nonconceptual content can explain a decrease in the perspective-dependence of those abilities of the system by reference to which the system's contents are specified. This explains how a cognitive science which models in terms of nonconceptual content can nevertheless satisfy conceptual constraints on cognition. I suggest that the interesting cognitive employment of connectionism should not rest on the S/S theory, because the S/S theory entails conceptualist theorising, and connectionist cognitive modelling is suited to nonconceptualist psychological modelling. I give some reason to think that C3 is as suited to (a way of cognitively using) connectionism as LOT is to classical AI. The potential for Connectionism to use nonconceptual content shows why Fodor and Pylyshyn's (1988) criticism of connectionism is misplaced. Connectionism can use the apparatus I have introduced to show how connectionist cognitive modelling can, in principle, respond to the problem of embodied cognition.
(2) The Problem of Embodied Cognition and the Construction Constraint on its Solution
2.1 The Problem of Embodied Cognition
Consider the following way to bring out the problem of embodied cognition.
Suppose, for the purposes of this article, that there is an irreducible and indispensable scientific level of cognitive explanation of human behaviour and that, even by the end of the next millennium, cognitive science will not have been made redundant by neurophysiology, quantum mechanics or some other noncognitive level of explanation.
Let us also accept naturalism: that all nonphysical properties are either reducible to or must be realised  or implemented in physical properties. In other words, anything that has a causal power either has only physical causal powers or must be built out of physical components, so that it is possible, in principle, to understand why it is that something which is built physically like that (pointing to the physical science description), has those causal powers (pointing to the nonphysical description). Naturalism does not require that nonphysical properties be - despite appearances - really physical properties (naturalism does not require reduction), but it does require that if we knew all the science there could be, we should not find it coincidental that certain physical objects have the irreducible nonphysical properties that they have.
Whether or not human behaviour can be explained physiologically, humans behave as they do because of neurophysiological properties of humans. But - given our first supposition - humans behave as they do because of certain irreducible cognitive properties. Neurophysiological explanation and cognitive explanation are independent of one another, and—apart from cognitive or physiological breakdowns—are each complete, in their own terms. How then can it be that both of the following are true: (1) cognitive explanations of behaviour are not causally redundant, and (2), the physical causation of the behaviour of a person marches in step with the cognitive causation of the behaviour of a person, so that a person is not torn apart in a tug of war between the physical and cognitive causal powers in the person. I write as I do because of my beliefs about how best to communicate a philosophical problem to a readership that is partly nonphilosophical, but it is also true that I behave as I do because of certain neurophysiological causes within me. How do we avoid the conclusion that there is a battle for the control of my hand?
This problem is not resolved by supposing that cognitive explanation is non-causal, for the problem will then re-emerge as: how is it possible that the behaviour of a person, which is physically caused, is coherent from a cognitive perspective? My writing as I am is cognitively predictable (whether or not it is cognitively caused), but had my neurophysiology been different in any of a very large number of imaginable ways, I would not be writing at all; because, for example, my hand would be motionless, or stuck behind my back. How can my physiology keep on making my body do one of the limited range of things that it must do if it is to make cognitive sense?
In short, how can we understand cognition naturalistically without either the reduction of cognitive properties to noncognitive properties, or the elimination of cognitive properties, or the rejection of the scientific indispensability of cognitive properties? To understand this is to understand how cognition can be physically embodied, and thus to understand how to solve the problem of embodied cognition.
2.2 The Construction Constraint
There is a naturalistic alternative to reduction, elimination and explanatory dispensability: the construction of cognitive properties out of noncognitive properties. This idea may be introduced by an example.
The notion of architectural functionality may be essential to the work of an architect, even though the notion cannot be reduced to any builder's notion of the arrangement of materials. For example, an architect may need to work with the notion of an efficient corporate headquarters. But this notion cannot be defined in terms of the spatial arrangement of commercial sizes of bricks, stone, metal, glass, plastic, wood and concrete. An unspecifiable infinite set of different arrangements of builders' materials will be sufficient for an efficient corporate headquarters, given a particular company at a particular stage of development and a particular technological and ethnographic context. Not only does an unspecifiable infinite set make reduction impossible, but which unspecifiable set this is will vary with the contextual parameters. What is efficient for a small company may not be efficient for IBM in the eighties. And what is efficient given telephones, electronic mail and fax machines would not be efficient in a technological context which predated these communication media.
So, the notions in terms of which a client would specify a building to an architect cannot be reduced to notions that a builder must work with. There are, thus, two distinct levels of notions (levels of description) which an architect must somehow bridge if he is to do his job. For a layman, the architect's ability to produce a builder's specification from an architectural specification, or to know which architectural properties would be instantiated by a building built to a builder's specification, may appear unintelligible. There is no further level of description that the architect employs. Rather, in learning his job, an architect has gained an understanding of the architectural notions and the builders' notions that allows him to move back and forth between descriptions at each of these levels. For the architect the relation between the two levels of description is Intelligible, not coincidental, whereas for the layman the relation is not Intelligible and so may appear coincidental.
The architect's understanding may be more practical than theoretical. Finding Intelligible the gap between the architectural level of description, and the materials level of description may consist simply in the following skill: given any building specification, an architect should be able to tell (and to know that he can tell) what architectural properties a building constructed like that would have, and given any architectural specification (eg. to provide functional office accommodation which is appropriate to the context of St. Pauls Cathedral), an architect should know how to put together building materials so as to satisfy the specification. For the skilled architect, the gap between the two levels of description is Intelligible not coincidental: we may say that the architect — but not the layman — has the ability to construct architectural notions from building notions.
Thus the relation of construction is an explanatory relation between levels which differs from the relation of reduction and from the relations of elimination and dispensability. The construction constraint, applied generally, claims that any non-physical level of description and explanation should be constructable out of, ultimately, some physical level, in an analogous sense to the architect's construction of architectural notions out of building and materials notions. If we have to construct some notion j out of physical notions then we need to be able to understand the nature of an object's being j in terms of a sequence of levels of description which are such that the top level makes manifest the j-ness of the object (as, for example, the architectural level makes manifest the corporate efficiency of a building), the bottom level is a physical level of description, and every two adjacent levels are such that the gap between them is Intelligible (not coincidental), as is the gap — for the architect — between the architectural and builders' levels.
Relying on this intuitive idea of the distinction between Intelligible gaps between levels and gaps which are coincidental or miraculous, the construction constraint can be stated more formally as follows:
A theory (or, rather, a framework of theories) of an ability j meets the construction constraint with respect to j if, and only if, it explains what it is for an organism to possess the ability in terms of the possession of a sequence of more than two levels of abilities, L1 - Ln, such that:
(i) the base level, L1, abilities are such that we do not understand why it is that an organism which has these abilities is thereby an organism which has the ability j, (that is, L1 and Ln are so related as to generate a miraculous coincidence problem about how they march in step), and,
(ii) the theory shows why it is that possession of the top level, Ln, abilities constitutes or manifests possession of j, and,
(iii) for each pair of levels, Li and Li+1 (1≤i<n), Li and Li+1 are not related in such a way that they generate a miraculous coincidence problem about how they march in step (that is, the gap between Li and Li+1 is Intelligible).
The gap between the folk-psychological level and the neurophysiological level is not Intelligible in this way: I exploited this gap to make vivid the problem of embodied cognition as a tug-of-war for the control of my hand. Given only our folk-psychological and neurophysiological knowledge, the marching in step of these two levels appears to be a miraculous coincidence. In contrast, because we know how to build machine languages out of electronic components and a high level language like LISP out of a machine language, the fact that the behaviour of the computer is determined by two levels of description (LISP and electronics) each complete in its own terms, does not appear as a coincidence. The construction relation between LISP and electronics does not require that there be law-like relations between any two adjacent levels: there are no known laws connecting electronic and computational descriptions. It requires just that somebody with knowledge of, for example, machine languages, LISP and electronics, should know roughly how to go about putting together electronics so as to build a LISP machine. Not that one has to succeed every time; just that one shouldn't find it coincidental that what has been built functions as a computer.
There is an interpretation of cognitive science according to which it has been exploring a construction of conceptual abilities in terms of interposing representational and computational levels between the folk-psychological and neuro-physiological levels. In the next section I show how the LOT theory can be seen as offering an account along these lines. The rest of the paper develops an alternative account which is more appropriate to the use of connectionist computational architectures and is not subject to the same kind of difficulties as LOT.
(3) The Language of Thought as Concept Construction
Remarkably, there has only been one serious attempt to solve the problem of embodied cognition by satisfying the construction constraint on the possession of concepts. This is the model which Fodor has called the Language of Thought (LOT), and he and others have defended at a philosophical level for over a decade. The possibility of LOT turns crucially on the possibility of both the computational and the psychological application of a theory of representation which has been developed in logic since Frege. This is the theory which characterises a representational system in terms of its combinatorial syntax and combinatorial semantics.
The syntactic theory of a representational system provides a recursive specification of all and only the legal concatenations of the atomic representations of the system. The semantic theory of a representational system provides an axiomatisable, recursive specification of the interpretation of all the legal representations. If the representational system is to do any work it must be possible to define over the syntax either a proof theory for logical work, or a theory of procedural consequence for computational work, which specifies all of the legal transformations from each legal representation. Much of the value of the logical tradition has rested on our being able to define purely syntactically a theory of legal transformation which nevertheless respects semantic constraints.
The syntactic and semantic theories must be explanatorily independent, yet linked. They must be independent in that it must be possible to understand how to apply the theory of legal transformation without understanding anything of the semantic theory; but they must be linked in that the syntactic application of the theory of legal transformation must not violate semantic constraints: traditionally, it must not transform a set of true premises into a false conclusion. And, the syntactic and semantic theories must be so related that even if the representational system is not complete, a useful proportion of all of the semantically coherent transformations must be capturable by the syntactic application of the theory of legal transformation.
What is so remarkable about LOT is its insight that the way to achieve the required Intelligible connection between the computational component and the psychological component of modelling in cognitive science is by developing a syntactic and semantic representational system for which the syntax is implemented computationally and the semantics is appropriate for psychological explanation. If this can be done, then the way to achieve the required relation between the computation and the psychology will simply follow from our understanding how to establish representational systems for which it is possible to specify a syntactic theory of legal transformation which respects semantic constraints. What the S/S theory delivers is a semantic - independent level (syntax) that marches in step with the semantic level. So if the syntax can be implemented computationally and the semantics can provide the basis for psychological explanation, then LOT will have shown how computational transitions can march in step with psychological transitions.
The syntactic characterisation of classical computational architectures is natural in part because these architectures themselves grew out of the logical tradition. And the psychological employment of semantics is natural because of a tradition that also goes back to Frege; the tradition of taking the meaning of a sentence to be the object of the propositional attitude which is expressed by the concatenation of a psychological verb (“believes", “desires", ...), “that" and the sentence itself.
Cognitive science, as interpreted by LOT, becomes distinct from both the theory of computation and the theory of psychology, because it is the attempt to establish empirically an intermediate level of explanation built on a system of cognitive representation which is such that, (1) the syntax is computationally implementable, (2), the semantics captures the important psychological generalisations, and (3), the theory of procedural consequence is consistent and usefully complete. It is of course true that work in cognitive science will proceed without LOT, that many cognitive scientists disagree with LOT, and that numerous objections have been raised to LOT. But LOT remains the only theory which gives clear criteria for the evaluation of the success of cognitive science, which is plausibly workable, and which shows how cognitive science may achieve the significance it hopes for: to explain how cognition can be physically embodied by constructing the psychological component of a cognitive science model out of the computational component.
My purpose here is not to criticise LOT, but to understand how to develop an alternative theory which has a comparable explanatory scope. LOT is so impressive because it rests on the remarkable tradition of S/S representational theory. To develop an alternative solution to the problem of embodied cognition, we need to develop an alternative theory of representation. An alternative theory of computation is not sufficient.
(4) Content, Conceptual Content and Nonconceptual Content
I have suggested how it is, in the case of LOT, that providing the kind of explanation appropriate to the problem of embodied cognition depends on the theory of representation which LOT employs. But it is also true that the choice of a representational theory determines the kind of psychological explanation which a model can offer. This is so because the representational theory determines the kind of content which can be assigned to states of the model, and this, in turn, determines the kind of psychological explanation that the model can make available. The link between representation and psychological explanation is content.
4.1 Introducing Content
I will begin with a pocket account of the notions of content, conceptual content and nonconceptual content, before presenting a more careful analysis of them.
Human persons act as they do, and thus often behave as they do, because some aspect of the world is presented to them in some manner. The term “content", as I shall use it, refers, in the first instance, to the way in which some aspect of the world is presented to a subject; the way in which an object or property or state of affairs is given in, or presented to, experience or thought. For example, I see the grey, plastic rectangular object in front of me as being a typing board, having the familiar Qwerty structure. I also see it as being in front of me, and these facts are responsible for my hands moving in a certain way. Representational states of mine have content in virtue of which they make the world accessible to me, guide my action, and (usually) are presented to me as something which is either correct or incorrect. I shall speak of a representational state (or vehicle) having content. It may be that a single representational vehicle carries more than one content, even more than one kind of content.
The theory of content - in terms of which we explain what content is - locates the notion with respect to our notions of experience, thought and the world. But it is important to see that this is consistent with the notion of content being applied to (though not explained in terms of) states which are not states of an experiencing subject. There are derivative uses of the notion in application to the communicative products of cognition, such as speech, writing, and other sign-systems, or to non-conscious states of persons such as sub-personal information processing states, but these uses must ultimately be explained in terms of a theory of the primary application of content in cognitive experience.
Conceptual content is content which presents the world to a subject as the objective, human world about which one can form true or false judgements. If there are other kinds of content, kinds of nonconceptual content, then that will be because there are ways in which the world can be presented to a subject of experience which do not make the objective, human world accessible to the subject. It is not unnatural to suppose that there must be nonconceptual forms of content, because this is the kind of thing that we want to say about very young human infants (before the acquisition of the object concept, say), or very senile people, or certain other animals. It is compelling to think of these beings as having experience, yet they are unable to communicate thoughts to us; we are unable to understand — from the inside — how they are responding to the world; we are unable to impose our world on them.
Conceptual content presents the world to a subject as divided up into objects, properties and situations: the components of truth conditions. For example, my complex conceptual content (thought) that the old city wall is shrouded in mist today presents the world to me as being such that the state of affairs of the old city wall being shrouded in mist obtains today. To understand this content I have to think of the world as consisting of the object, the old city wall, the property of being shrouded in mist, and the former satisfying the latter. The possession of any content will involve carving up the world in one way or another. There will be a notion of nonconceptual content if experience provides a way of carving up the world which is not a way of carving it up into objects, properties or situations (ie. the components of truth conditions).
It is natural to say that the possession of content consists in having a conception of the world as being such and such. But the word “conception" is too closely related to “concept" for it to function neutrally as between conceptual and nonconceptual presentations of the world. I shall say that a content registers the world as being some way, and so ask, is there a way of registering the world which does not register it into objects, properties or situations?
4.2 Definitions of Conceptual and Nonconceptual Properties
I will begin a more careful analysis of these notions by introducing definitions of conceptual and nonconceptual properties, and then show how these definitions can be applied within the theory of content.
A property is a conceptual property if, and only if, it is canonically characterised, relative to a theory, only by means of concepts which are such that an organism must have those concepts in order to satisfy the property.
A property is a nonconceptual property if, and only if, it is canonically characterised, relative to a theory, by means of concepts which are such that an organism need not have those concepts in order to satisfy the property.
Notice that the difference between these two definitions lies principally in the difference between the italicised “must have" in the first definition, and “need not have" in the second definition.
Consider the property of thinking of someone as a bachelor. A specification of what this property is will use the concepts *male*, *adult* and *unmarried*. But nothing could satisfy the property unless it possessed these concepts, since nothing would count as thinking of someone as a bachelor, unless he or she was able to think of the person as being male, adult and unmarried. So the property of thinking of someone as a bachelor (unlike the property of being a bachelor) is a conceptual property.
Or consider the belief property of believing that the Stanford Campus is near here (where I think of the Stanford Campus as the Stanford Campus, rather than as the campus of the richest university in the West, and I think of here as here, rather than as 3333 Coyote Hill Road). Given this, nothing could satisfy the property unless it possessed the concept of the Stanford Campus qua Stanford Campus. Thus the property is canonically characterised only by means of concepts which an organism must have in order to satisfy the property, and is therefore a conceptual property. Contrast the property of having an active hypothalamus. Such a property is characterised by means of the concept *hypothalamus*, but an organism may satisfy the property without possessing this concept. Therefore the property of having an active hypothalamus is a nonconceptual property.
Formally, the idea is that conceptual content is content which consists of conceptual properties, while nonconceptual content is content which consists of nonconceptual properties. Can we give any substance to this formal idea?
4.3 The Application of the Definitions of Conceptual and Nonconceptual Properties within the Theory of Content
In order to show that there is a notion of nonconceptual content we need to show that the definition of nonconceptual properties can be applied within the theory of content. What does this mean?
The definitions of conceptual and nonconceptual properties use the notion of canonical specification, for otherwise every property would be a nonconceptual property, since, trivially, every property — including conceptual properties — can be specified by means of concepts that the subject need not possess. So we need to employ the notion of canonical specification. If we are to apply these definitions within the theory of content then the notion of canonicality that we are interested in is the notion of being a canonical specification within the theory of content. Certain specifications of a state or an activity are identified within a theory of content as being canonical when they are specifications generated by the theory in order to capture the distinctive way in which some aspect of the world is given to the subject of the state or activity. So, as brought out by McDowell (1977), “'aphla' refers to aphla" would be canonical, but “'aphla' refers to ateb" would not be, even though both would be true, because aphla is ateb. The notion of being canonical within the theory of content is parallel to the notion of being canonical in the theory of number, where the canonical specification of the number nine is not “the number of planets", but “nine".
4.31 The Case of Conceptual Content
4.311 The Notion of a Task Domain
In order to understand how conceptual content works we need the notion of a task domain for a behaviour. A task domain is a bounded domain of the world which is taken as already registered into a given organisation of a set of objects, properties or situations, which contains no privileged point or points of view, and with respect to which the behaviour is to be evaluated.
SHRDLU's blocks micro-world was SHRDLU's task domain. The notion of a Model in formal semantics, and (often) the notion of a possible world in logic are notions of task domains. Likewise, the performance of a chess computer is evaluated with respect to a chess task domain which consists of 64 squares categorised into two types, 32 pieces — each with an ownership property —, a legal starting position, three types of legal ending position, and a set of transformations from each legal position to all of the legal continuations from that position. The computer's task domain excludes, for example, human emotions and plans, lighting conditions, reasons for, and the point of, winning ... What this means is that the performance of a chess playing computer is evaluated with respect to transformations of chess tokens on a 64-square board, but not with respect to its response to human emotions, the lighting conditions, the historical pattern of the game, or “its reasons for winning". Moreover, because the domain is fixed so that certain situations are registered as wins for White, and certain others as wins for Black, the performance of the computer is not assessed with respect to its ability to transfer its knowledge to a different game, chess*, which is identical to chess except that those situations which are wins for White in chess, are wins for Black in chess*, and those situations which are wins for Black in chess are wins for White in chess*.
A task-domain, then, is a conceptualised region of the world which provides the context of evaluation (true / false, win / lose, true-in-a-model / false-in-a-model, adaptive / non-adaptive, successful / unsuccessful ...) for the performance of some system. How is the notion of a task-domain connected to the notion of conceptual content?
4.312 The Specification of aContent by Concepts of the Task Domain
Consider again the cognitive occurrence in me that we express in words as, “I am thinking that the Stanford Campus is near here". This is a representational state of mine, and may possess more than one kind of content. What kind of content does the state carry? There is a type of content (let us call it, “a content") which is stipulated within the theory of content to be a kind of content that has determinate truth conditions; that is, whose evaluation as correct imposes a determinate condition on the world. It follows that the linguistic expression, “that the Stanford Campus is near here" cannot fully capture the a content of the representational state, since this requires a fixed interpretation for “near" and “here". (In order for the state to be a state with a content, we need to know what truth condition it imposes on the world. But the words “here" and “near" do not tell us.)
Now suppose that this state occurs as part of a project of mine in which I am planning how best to eat lunch given various parameters and constraints on me: time, money, hunger, distance to eating locations, speed of transport available to me, cost of food at various locations. These parameters and constraints establish a task domain which fixes an interpretation for the terms “near" and “here": suppose that it follows from the time constraints on me, and my hunger, that I need to be eating within fifteen minutes. Then “near" means: can be reached by a mode of transport available to me within fifteen minutes. Likewise “here" will mean something like: the region between the spot on which I am standing and a line joining the embarkation points for all the modes of transport which are part of my planning domain.
The interpretation of my cognitive occurrence as having a content depends on specifying the content by means of concepts of a task domain; in this case, the domain of my planning to eat lunch under various constraints and given various parameters. In other words, the provision of determinate truth conditions for my cognitive state, required by the interpretation of it as having a content, entails that the content is canonically specified by means of concepts which reflect the objective structure of the task domain: its organisation into objects, properties and situations. Since an organism can only grasp an a content if it grasps its truth conditions (or its contribution to the truth conditions of contents containing it), it follows that an organism which grasps such a content must know what the (relevant part of the) t-domain of the content is. But a t-domain (unlike the world) is essentially conceptually structured, so there is no way of knowing what the t-domain of a content is without possessing the concepts in terms of which the t-domain is structured. Hence possession of an a content requires possession of the concepts in terms of which it is canonically specified. It follows that a content is a kind of content which consists of conceptual properties, as defined above. That is, a content is conceptual content.
The process of identification of a content as conceptual content may be mimicked in order to demonstrate a notion of nonconceptual content. We must ask, is there a way to motivate in a similar fashion the application of the definition of nonconceptual properties within the theory of content? In asking this, I am asking whether nonconceptual specifications of states or activities can ever be canonical within the theory of content. Thus I am asking whether nonconceptual specifications of an activity can ever be required by a correct theory of content in order to capture the distinctive way in which some aspect of the world is given to the subject of the activity.
We can clarify what is involved in doing this by setting out, as a summary of the above discussion, the different elements that I have used in motivating the definition of conceptual properties within the theory of content:
(1) The definition of conceptual properties (by stipulation);
(2) The claims that there is a constraint within the theory of content which requires determinate truth conditions, and that possession of content which satisfies this constraint requires knowledge (grasp) of its truth conditions (these claims are given by the theory of content, and are constitutive of this notion of content);
(3) A psychological state expressed linguistically as “thinking that the Stanford Campus is near here", not yet analysed with respect to the kind of content that it has;
(4) The claim, argued in the text, that the interpretation of (3) under (2) requires the notion of a task domain and the specification of the content of (3) by means of concepts of the task domain.
(5) (4) results in the satisfaction of (1), hence the identification of content which satisfies the constraint in (2) as conceptual content.
The notion of a task domain provides the link between the philosophical notion of a content, and my stipulative definition of conceptual properties; a link which is needed to show that the analysis of a psychological state in terms of a content entails satisfaction of the definition of conceptual properties.
4.32 The Case of Nonconceptual Content
I can show the need for nonconceptual content by showing that there are psychological states the full understanding of which requires a notion of content which cannot be analysed in this way; that is, which must be canonically specified by means of concepts that the subject need not have. The discussion will have to parallel the discussion for the case of conceptual content, so we need a parallel for (1) - (5):
(1') The definition of nonconceptual properties (by stipulation);
(2') Some constitutive conditions on a kind of content, b content, which are provided by the theory of content, but which are different from the conditions in (2).
(3') Some psychological or representational state as yet unanalysed with respect to the kind of content it has.
(4') An argument for the claim that the interpretation of (3') under (2') requires the notion of some domain other than the task domain and the specification of the content of (3') by means of concepts of this domain.
(5') A demonstration that (4') results in satisfaction of (1'), hence the identification of b content as nonconceptual content.
We already have (1'). What about (2')?
4.321 Cognitive Significance
A good theory of content is answerable to various constraints. For example, a good theory of content should be appropriate for use within a content-based scientific psychology, it should have resources to explain how certain contents have determinate truth conditions and a good theory of content should also capture cognitive significance, that is, the role that content plays with respect to perception, judgement and action.
How can the theory of content accommodate cognitive significance? Frege's notion of sense was introduced, in the first instance, to explain how certain identity statements could be informative. For example, to learn that Hesperus = Hesperus is not to learn anything new, but to learn that Hesperus = Phosphorus may be to learn something of considerable significance, yet Hesperus is Phosphorus. It follows that possession of the content expressed here by the word “Phosphorus" cannot consist just in the ability to think of the planet Venus, (specified no further than this), because just the same ability is associated with “Hesperus". There is here a motivation for introducing a notion of content (sense) which differs from a purely referential notion of content (reference). There is a content expressed by “Hesperus" which is different from the content expressed by “Phosphorus" because the former content plays a different role from the latter content in a person's judgements of the truth value of contents of the form “... = Hesperus". Frege generalised this motivation into a criterion of identity for such contents (senses). We may generalise it still further to yield a generalised notion of sense which I call “b content", whose identity conditions are fixed, not just by its constitutive connections to judgement, but by its constitutive connections to perception, action and judgement. Possession of a particular b content requires possession of a contentful state which plays that role in the psychological economy of the subject which is constitutive of the b content.
A major success within recent work in the theory of content has been to show that there are indexical and demonstrative b contents that cannot be canonically specified, in the way appropriate to conceptual content, by means of any description. This has been achieved by showing that were a description — per impossibile — to provide canonical specification of the content, in the way appropriate to conceptual content, it would alter the cognitive significance of the content, that is, the character of its constitutive connections to action and judgement. Since cognitive significance is constitutive of b content, it follows that this form of specification cannot canonically capture b contents.
For example, Perry (1979) shows this for the indexical “I" and connections to action, and Peacocke (1986) shows it for demonstrative perceptual contents and connections to perception and judgement. Perry's point is that the conceptual use of any descriptive canonical specification — *the x such that fx* — for the indexical content *I*, will alter the cognitive significance of the thought *I am y* by altering its constitutive connections to action. The reason for this is that it is always possible that one may not realise that I am the x such that fx, so that even if one would act immediately on the basis of judging *I am y* (eg, *I am spilling sugar all over the supermarket floor*), one might not act on the basis of judging *the x such that fx is y*.
Peacocke contrasts what a person knows when he or she knows the length of a wall in virtue of just having read an estate agent's handout, and what a person knows when he or she knows the length of a wall just in virtue of looking at it. Frege's intuitive criterion of difference for contents can be used to show that although both people know the length of the wall, neither knows what the other knows. Thus suppose that my wife's and my cognitive states were identical except for the fact that I know what the length of the wall is just in virtue of having read the handout, and she knows what the length of the wall is just in virtue of having seen it. But then, thinking of the length of the wall in only that way which is available to each of us, I may be agnostic about the thought *that length is greater than the length of our piano* (because, for example, we don't know how long in feet our piano is), whereas my wife will judge this thought to be true because, simply by looking, she can see that our piano will fit against the wall. Therefore, the perceptual demonstrative b content differs from any descriptively specified conceptual content, and so cannot be canonically specified, in the way appropriate to conceptual contents, by means of any specification such as, “the person sees that distance-in-feet(a,b) = n" where a and b are the end-points of the wall.
We could treat examples such as Perry's and Peacocke's in a way which was similar to my treatment of thinking that the Stanford Campus is near here — that is conceptually — by means of concepts of the respective task domains. That would be, in effect, to characterise these indexical contents in a descriptive, conceptual fashion. But Perry's and Peacocke's arguments show that justice cannot be done, in such a way, to the cognitive significance of these contents. So we have only to recognise a notion of content for which cognitive significance is essential, to see that there is a kind of content which cannot be canonically specified by means of concepts of the task domain.
The argument so far shows that there is a very large class of cognitive states (all states which contain indexical or demonstrative elements) which have a kind of content (b content) for which the only canonical conceptual specification is the use of a simple demonstrative or indexical under the conditions of a shared perceptual environment or shared memory experience. Such a specification is evidently useless for the construction-theoretic purposes of a scientific psychology since the only way the theorist can have to understand the nature of the content is either to share the experiential environment of the content, or draw on similar experiential environments available to the theorist in memory experience. (Scientific psychology, here, is psychology which is aiming to solve the problem of embodied cognition, and which therefore is aiming to construct any explanatorily indispensable notion of content out of non-content involving levels of description). Yet this class of contents is particularly important for psychology, at least because of its direct connections to action and its crucial role in learning. Is the theoretical psychologist therefore incapable of capturing those contents which are basic to our ability to act in the world and to learn from it?
Only if the psychologist assumes that he or she must work with conceptual content. The problem arises because there is no conceptual structure within the demonstrative or the indexical or the observational content which can be exploited to yield a canonical conceptual specification of the content which would be appropriate for the purposes of a scientific psychology. But this doesn't exclude there being any nonconceptual structure within the content. If we can make sense of this notion, then there is here an argument to show that much of the psychological life can only be captured by means of, and should, therefore, only be modelled in terms of, nonconceptual content.
4.322 The Notion of a Substrate Domain
Abandon, then, the demand that every content must have its theoretical specification given in the way which is constitutive of conceptual content; that is, by means of concepts of the task domain. What other theoretically adequate method of specification could there be? I introduce below one kind of canonical nonconceptual specification. It is not necessarily the only kind, although I believe that it is the only kind in terms of which we can solve the problem of embodied cognition.
It will help to consider the operation of an autonomous, mobile robot known as “Flakey" which lives at a research institute, SRI, in California. Flakey navigates the corridors of SRI. His task is to move up and down the corridors, avoiding hitting the walls, and to turn into particular doorways.
In order to be able to behave flexibly in a range of task domains a system must be able to employ representations of features which are special to the domain in which it happens to find itself. For example, if the width of corridors varies in Flakey's environment, then Flakey will need to respond differentially to corridor width. Given the kind of system that Flakey is, this will mean that Flakey will have to represent this variable. The system need only not represent that which does not change throughout the career of the system. So the greater the system's representational capacity, the greater its potential flexibility. Should we suppose, therefore, that the cognitively ideal system would computationally represent — in the traditional AI style — all the facts there are? That although nothing achieves this ideal, the closer one comes to it, the better one's cognitive capacities will be?
To suppose this is to miss an important distinction between two kinds of fact. What I want to show is that computational representation of only one of these kinds of fact is required for the ideal Artificial Intelligence system. Flakey is sometimes imagined to deliver pizza throughout SRI. It might be that only one weight of pizza is allowed through the extensive security system, and that Flakey could therefore be built on the assumption that if something is recognised as a pizza, then the mobile arm needs to exert a certain force to lift it. This would have the effect of “unburdening" the representational capacities of Flakey, with respect to having to work out each time it was about to lift a pizza, how much force was required to lift it. This connection could simply be built into the hardware. However, the folks down at Hewlett Packard, intrigued by Flakey's growing reputation, might want to try him out on delivering pizza for them. They would be sorely disappointed because, unfortunately for Flakey, the security system at HP labs lets all weights of pizza through. Flakey was discovered to be throwing pizza around in a way not likely to impress DARPA.
Indeed, DARPA could reasonably argue that this was a cognitive defect of Flakey's. We treat intelligence in an open-ended way: So-and-so may be great at chess, but if he can't learn to play Go, then we think him the less intelligent for it. For Flakey, representation of pizza weight is required for acceptable, let alone ideal, cognition.
But we shouldn't conclude therefore that to be truly intelligent Flakey must represent all the facts there are. For example, it would be surprising if Flakey were to represent the distance between the sonar sensors at its base. This is not only for the reason that this distance is a constant throughout Flakey's career, but, more importantly, because Flakey's own structure is not part of Flakey's task domain. Flakey never has to manipulate the distance between his sonar sensors; this distance is not something with respect to which Flakey's performance will be evaluated. Rather, it is part of Flakey's substrate of abilities in virtue of which Flakey has those corridor-movement behavioural capacities which he in fact has. This distinction between task-domain (“t-domain") and the domain of the system's substrate of abilities (“s-domain") is essential to understanding what a flexible system is required to represent. To be able to operate flexibly in a range of t-domains a system must be able to represent those features of a t-domain which vary, or may vary, within the range of t-domains. But so long as the s-domain is outside this range, as it usually will be, a flexible system has no need to represent aspects of its s-domain.
My visual capacity may be quite superb, and open-ended: I can visually discriminate any kind of object, in an extensive range of conditions of illumination, and distances from me, and so forth. But nobody would suggest that it is a defect of my visual capacity, that I am ignorant of the algorithms employed by my visual information processing system. With respect to my personal level visual capacity, my sub-personal information processing capacities are part of the s-domain. Given a division between t-domain and s-domain in a particular case, performance in the task domain— even fully conceptual performance — does not require the possession of any concepts of the s-domain.
4.323 Specifying bContents by Concepts of the Substrate Domain
As we saw, the notion of a t-domain provided the link between a content and the definition of conceptual properties. Can the notion of the s-domain provide a parallel link between b content and the definition of nonconceptual properties? An intelligent agent does not need to have concepts of its s-domain, so if b content can be canonically specified by reference to the objects and properties of the s-domain, we will have motivated a kind of content which is specified by means of concepts that the system or organism need not have.
Consider the following quotation from Evans (1982, chapter 6):
What is involved in a subject's hearing a sound as coming from such and such a position in space? ... When we hear a sound as coming from a certain direction, we do not have to think or calculate which way to turn our heads (say) in order to look for the source of the sound. If we did have to do so, then it ought to be possible for two people to hear a sound as coming from the same direction and yet to be disposed to do quite different things in reacting to the sound, because of differences in their calculations. Since this does not appear to make sense, we must say that having spatially significant perceptual information consists at least partly in being disposed to do various things.
When Evans asks, “what is involved in a subject's hearing a sound as coming from such and such a position in space?", he is asking about the nature of the content by which the subject is presented in experience with this aspect of the world. Evidently the content is indexical or demonstrative since, were we to express the content in words, we would say that perception presents the sound as coming from “that location", or “from over there". The conclusion drawn on the basis of Perry's and Peacocke's examples applies: there is no way to canonically specify this content as a conceptual content, if we wish to do theoretical justice to the cognitive significance of the content; in particular its direct connection to action. What Evans adds, is, first, a further reason why this kind of content cannot be captured conceptually (no conceptual content can be necessarily linked to action as directly as certain b contents require), and, secondly, the suggestion that the way to capture the cognitive significance of the content is by reference to a way of moving in the world; the subject's ability to reach out and locate the object, or walk to the source of the sound, which the perceptual experience makes available. At the place in the argument which we have now reached, it is this second idea which is important, because, for Evans's content, a way of moving in the world is part of the s-domain.
Given our usual views about consciousness, the idea here can seem quite strange: it is the idea that certain contents consist in a means of finding one's way in the world (tracking the object, say