Introduction Symbols and symbol manipulation have been central to cognitive science (Fodor, 1975; Fodor & Pylyshyn, 1988; Minsky, 1983; Newell & Simon, 1976). But the disembodied nature of traditional symbolic systems has been troubling many cognitive scientists (Agre, 1988; Bickhard, 1993; Churchland, 1986; Dreyfus & Dreyfus, 1987; Freeman, 1995; Searle, 1980; Sun, 1994; Varela et al., 1993; Waltz, 1990; Wertsch, 1991; Winograd & Flores, 1987). One remedy that has been proposed is symbol grounding, that is, connecting symbols to lower-level sensory-motor procedures and thus rooting the abstract in the concrete (Barsalou, 1999; Harnad, 1990).

In this paper, we will take a detailed look at this old issue. Aided by our new work on computational (in a broad sense) cognitive modeling, which we have previously published extensively on (Sun, 1997; Sun & Peterson, 1998a,b; Sun et aL, 1998a,b), we hope to further our understanding of this issue. We will do so by utilizing a concrete example of a model that offers a new perspective on matters related to this issue and beyond. The new ideas to be offered from this model include the hypothesis of dual processes (mediated and unmediated interaction), bottom-up learning (a two-stage process of learning symbolic representation), and symbol grounding in direct "comportment." Armed with our new perspective, we will also go back in time, to link up with traditional philosophical accounts relevant to this issue, such as Heidegger (1927a).

My main points can be summarized as follows. Symbols should be grounded, as has been argued for many years. But we insist that they should be grounded not only in subsymbolic activities (subsymbolic "representation"), but also in the direct interaction between the agent and the world. The point is that concepts, which symbols represent, are not formed in isolation (from the world), in abstraction, or "objectively." They are formed in relation to the life-world of agents, through the perceptual/motor apparatuses of agents, linked to their goals, needs, and actions. This view is argued on the basis of Heideggerian philosophy, which emphasizes the primacy of direct, unmediated interaction between agents and the world. Symbolic representation and concepts are derived from such direct interaction. Precisely in this sense can symbols really be grounded. We will show how this can be achieved computationally.

The remainder of this paper is a detailed account that extends from computational (in a broad sense) mechanisms to philosophical abstractions. In the next section, I will review the traditional notions of symbols, representation, and so on, and identify their important properties. In the third section, I will offer a theoretical perspective on these issues that remedies shortcomings of traditional approaches, drawing ideas from Heideggerian philosophy. In the fourth section, the mechanistic (computational) underpinning of this perspective will be analyzed and viable implementations suggested. In the fifth section, the framework thus far outlined will be analyzed in light of the issues raised in the first section. In the final section, further discussions will complete the paper.

Symbols and representation: some background

In this section, I will review some background notions needed in our discussion, including the notions of symbols and representation. I will try to clarify a few possible confusions and lay the foundation for later exposition of our new perspective (and our new models). In so doing, I will identify a few relevant properties of each of these notions, all of which I will utilize later on.

Symbols

Symbols have been a mainstay of cognitive science' ever since its early incarnations as information processing psychology and AI in the 1950s. The idea is based on the notion of computation as commonly understood in the early days (so in some sense, it is based on a simplistic and narrowly conceived notion of computation). Computation consists of input, output, retrieval, storage, and manipulation of symbols (such as in a von Neuman digital computer); a sequence of specified steps accomplishes a computational task, such as numerical calculation. Cognition is understood in much the same way, as consisting of input/output, storage/retrieval, and symbol manipulation. The most important part of this is, of course, the manipulation of symbols that changes symbol structures that stand for mental states.

Because of the use of this computer metaphor, cognition is perceived as a sequence of explicit and clear-cut steps involving nothing but symbols.

Later, the physical symbol system hypothesis introduced by Newell and Simon (1976) clearly articulated this "vision" and called for a concentrated research program along the symbol manipulation line. They claimed that "symbols lie at the root of intelligent action." They defined, as the fundamental building block of the science of the mind, physical symbol systems:

A physical symbol system consists of a set of entities, called symbols, which are physical patterns that can occur as components of another type of entity called an expression (symbol structure). Thus a symbol structure is composed of a number of instances (or tokens) of symbols related in some physical way (such as one token being next to another).

They further claimed that symbols could designate arbitrarily: "a symbol may be used to designate any expression whatsoever"; "it is not prescribed a priori what expressions it can designate." "There exist processes for creating any expression and for modifying any expression in arbitrary ways." Based on that, they concluded: "A physical symbol system has the necessary and sufficient means for general intelligent action," which is the famed physical symbol system hypothesis (where we take "general intelligent action" to mean the full range of human intelligent behavior; Newell & Simon, 1976). Clearly, a physical symbol system is an abstracted view of a digital computer (that is, it is an instance of a Turing machine, which is hypothesized in turn by Turing to be able to capture any "computational process"; see Turing, 1950). Now the loop is almost complete: if you believe in some kind of universality of "computation" (especially in cognition), and if you believe in Turing's hypothesis (the universality of Turing machines), then you are naturally inclined to believe in the physical symbol system hypothesis (because physical symbol systems, as defined by Newell & Simon, are Turing equivalent).

The physical symbol system hypothesis has spawned (and was used to justify) enormous research effort in traditional AI and cognitive science. This approach (classical cognitivism) typically uses discrete symbols as primitives, and performs symbol manipulation in a sequential and deliberative manner. Although this view came to dominance in AI and cognitive science for a while, it has been steadily receiving criticisms from various sources (for example, Bickhard, 1993; Dreyfus, 1972; Dreyfus & Dreyfus, 1987; Searle, 1980; Winograd & Flores, 1987). They focused on the disembodied abstractness of this approach. In response to such criticisms, Vera and Simon (1993) presented a modified version of physical symbol systems as follows: "A physical symbol system is built from a set of elements, called symbols, which may be formed into symbolic structures by means of a set of relations." "A physical symbol system interact with its environment in two ways: (1) it receives sensory stimuli from the environment that it converts into symbol structures in memory; and (2) it acts upon the environment in ways determined by symbol structures that it produces." It "has a set of information processes that form symbol structures as a function of sensory stimuli" as well as "produce symbol structures that cause motor actions." Clearly, in this new version, they tried to put more spins on sensory-motor connections, through which (they hope) symbols can be put in contact with the world. Can either of these two versions of the physical symbol system hypothesis justify the claim to universality? Can the new version sustain better the claim?

In this regard, let us look into a basic question: what are symbols after all? "Symbols are patterns," according to the new version of Vera and Simon (1993), in the sense that "pairs of them can be compared and pronounced alike or different." Patterns are symbols when "they can designate or denote." Then, the question becomes: what is the difference between symbols and, say, pictures?

According to the old version of the physical symbol system hypothesis, their answer would be that symbols, different from non-symbols such as pictures, can designate arbitrarily; according to the new version, however, their answer is likely to be that there is really no difference. The old version seems overly restrictive: why do we restrict ourselves to a particular type of pattern and forego the others? How can we believe that such a restricted type of pattern is necessary and sufficient for cognition? The new version seems overly liberal: if anything is a symbol, then of course cognition can be modeled by symbol manipulation; it becomes a tautology and thus trivially true.

In order to pinpoint precisely what a symbol is, we should abstract its essential characteristics. Those characteristics that are at issue here are the following: (1) arbitrariness: whether a pattern (or a sign) has an intrinsic meaning or not; and (2) syntacticity: whether a set of patterns or signs can be arbitrarily and systematically combined in a language-like manner (i.e. whether they have the compositionality and systematicity found in human languages). These two characteristics constitute necessary conditions for a pattern (or a sign) to be a symbol [1]. It is important to emphasize the distinction between the two different notions: signs (generic patterns) and symbols, which Peirce made clear a century ago (see Peirce, 1955). To quote from Peirce (1955):

A sign is either an icon, an index, or a symbol. An icon is a sign which would possess the character which renders it significant, even though its object has no existence; such as a lead-pencil streak as representing a geometrical line. An index is a sign which would, at once, lose the character which makes it a sign if its object were removed, but would not lose that character if there were no interpretant. Such, for instance, is a piece of mould with a bullet-hole in it as sign of a shot ... A symbol is a sign which would lose the character which renders it a sign if there were no interpretant. Such is any utterance of speech which signifies what it does only by virtue of its being understood to have that signification.

Because of the nonsensical consequence of the new version of "symbols," we will have to rely on the old version (Newell & Simon, 1976).

Whether symbols are necessary and sufficient for accounting for cognition is not a settled matter. However, almost nobody disputes that some form of symbols, in connectionist, classical, or some other ways, is needed for accounting for high-level cognition, for example, language and reasoning. Even radical connectionists accept that symbols may emerge from dynamic interaction of elements of a network (or a dynamic system in general). I do not make any claim here as to what form symbols should be in, but that there should be some for obvious reasons (Sun, 1994) [2].

Representation

Classical cognitivism believes that an agent has an internal copy, an explicit model of some sort, of the external world. In that model, there are internal states that encode external states in the world (Fodor & Pylyshyn, 1988). Cognition is accomplished by developing, maintaining, and modifying these internal states, i.e. representation (which is the basic tenet of representationalism). According to the analysis by Peirce (1955), a representational system consists of, or can be analyzed into, representational media, representational entities (i.e. what constitutes a representation), representational semantics (or references, i.e. what is being represented).

To understand characteristics of (explicit) representation (as in traditional cognitive science; Fodor, 1975; Collins & Smith, 1988), we can identify the following syntactic properties. First of all, such representation is explicit. This is because, without this requirement of explicitness, everything is a representation and thus the tenet of representationalism becomes meaningless. For example, a tennis ball has a representation of forces hitting it and trajectories it flies through, since it can respond to forces and fly through space; a car has a representation of roads and driving movements of its driver, since it can follow the driver's direction and stay on the road (Port & van Gelder, 1995). Thus, the representation thesis (representationalism) becomes trivially true in this way. Second, explicit representation is punctate: it consists of clearly delineatable items. Third, it is also elaborate: it contains much detail, even though it may not be a complete model [3]. Fourth, explicit representation, as has been traditionally used, is symbolic and compositional. Note that, although explicit representation need not be symbolic in the full sense of the term (exceptions include imagery, analogue, and so on), almost all the existing representational models and systems in traditional cognitive science were symbolic (see Anderson, 1983; Davis, 1990; Klahr et al., 1987; Minsky, 1983, for accounts of traditional systems) [4]. In addition, explicit representation is semantically specifiable: each particular representation can have a specific meaning as (arbitrarily) designated, and meanings are compositional as well.

There are many possibilities in terms of manipulating (explicit) representation. Symbol manipulation is the prime candidate for such a task. The compositionality of the syntax and semantics of explicit representation makes it easy to construct a computational procedure to manipulate representation and keep track of references and meanings. There are also other possibilities; for example, connectionist models (or neural) networks can be suitable for manipulating representation as well (see Abrahamsen, 1991; Bechtel & Clark, 1993; Chalmers, 1989; Sun & Bookman, 1994).

Intentionality

One of the most important questions concerning representation is the following: in virtue of what does representation have or acquire its meanings or signification? How does it come to represent what it purports to represent [5]? As we discussed earlier, it is doubtful that an arbitrarily constructed symbolic representation can suffice, by itself, to account for cognition, when only complex relations among symbols are involved (Newell & Simon, 1976). Clearly, meanings can hardly lie solely in symbols and their interrelations. How is it possible to move from the notion of meaning as arbitrary designation to a notion of "intrinsic" meaning? In other words, how do we relate representation to whatever is out there in the world, in an intrinsic way? No argument more clearly demonstrates this point than Searle's Chinese Room argument (Searle, 1980). The issue brought to light by this argument is intentionality. That is, our mental states and mental representation are about something. Mere symbols, as pointed out by Searle (1980), have no intentional content and thus cannot capture cognition adequately.

Cognitive science, especially AI, has been grappling with the issue of intentionality ever since the publication of Searle's (1980) argument [6]. Where may an answer to these questions lie? I would venture to suggest a few possible places that we may go to look for answers. These places include: (1) the existential experience of cognitive agents, individually or collectively, especially (but not exclusively) their everyday activities; and (2) their functional predispositions in such activities (which are acquired through evolutionary and other processes), including the biological substrates that embody their representational structures, functional capacities, and behavioral propensities [7]. Symbol grounding in the sense of linking symbols (symbolic representation) to lower-level (subsymbolic) processes (Hamad, 1990) provides a partial answer to the intentionality question. But it does not fully answer the question. This is because the question of how grounded symbols, and associated subsymbolic processes, acquire their intentional content remains. I would like to argue that, instead of being narrowly and technically conceived, symbol grounding should be understood in a broadened context in order to fully address the intentionality question, which is at the heart of the matter.

Let us look into some details of this position. Everyday activities and symbol grounding What is the structure of everyday experience of cognitive agents? How is representation acquired in that experience? Where do its meanings lie?

Let us draw some ideas from phenomenological philosophers such as Martin Heidegger (1927a,b) and Maurice Merleau-Ponty (1962, 1963) (and also Bickhard, 1993; Bruner, 1995; Dewey, 1958; Gibson, 1950, 1979; Rorty, 1979, 1991). What is particularly interesting is the notion of being-in-the-world (Dreyfus, 1992a,b; Heidegger, 1927a,b; King, 1964; Okrent, 1996). The idea is that our existence in the world, or our being-in-the-world, is fundamental to us, to our being what we are. Being-in-the-world entails that we constantly interact with the world in a natural, immediate, and non-reflective (i.e. reflexive) way, in our everyday activities in the world. It is believed that such "mindless" everyday activities or coping with the world (on top of our biological pre-endowment) is the basis of our high-level thinking and its intentionality [8].

According to Heidegger, our everyday coping with the world presupposes a background of common, everyday practices (Dreyfus, 1972, 1982, 1992a,b). The "background of practices" (Dreyfus, 1992a,b) is not represented in an explicit and elaborate fashion (such as what we see in traditional symbolic representation, e.g. in the rule base of an expert system; cf. Anderson, 1983; Klahr et al., 1987), which spells out every detail and every minute variation of every possible situation (Carnap, 1969), but is assumed in our comportment toward the world. In other words, the most important and fundamental part of our mind is embodied and embedded, not explicitly represented, and thus it is not directly and explicitly accessible to our (critical) reflection.

Another important theoretical notion is behavioral structure (or form; MerleauPonty, 1963; Madison, 1981). Maurice Merleau-Ponty extended Heidegger's notions, emphasizing the important role of the structural whole in the understanding of agents. It is not just external situations and internal reactions, but the structural connection that links the two that matters (the affordance and the effectivity, in Gibsonian parlance; Turvey, 1992). Situations and reactions are linked on the basis of their shared participation in structures that are comprised of agents and the world (for further discussions, see also Agre & Horswill, 1997; Ballard, 1991; Hammond et aL, 1995; Hutchins, 1995; Zhang & Norman, 1994) [9]. According to MerleauPonty, in such structures there lies the key to the understanding of meanings or signification of agent's behavior.

Comportment

Heidegger (1927a,b) proposed that there is a primordial kind of comportment in an agent that directly involves the agent with the world, without the mediation of (explicit) representation. It is a pattern of direct interaction between the agent and its environment, the world. As he put it, "comportment has the structure of directing-oneselves-toward, of being-directed-toward" (Heidegger, 1927b). This term is meant to capture the direct two-way interaction of an agent and its world (the "dialectics"; Merleau-Ponty, 1963).

Let us explore this notion further to better understand the interaction and the mutual dependency between an agent and its world (at the subconceptual level; Smolensky, 1988). First of all, comportment is direct and unmediated. Thus it is free from representationalist baggage. In other words, comportment does not necessarily involve or presuppose (explicit) representation and all the problems and issues associated with explicit representation (as discussed earlier). To the contrary, (explicit) representation, and relations between mental states and their objects, presuppose the existence of comportment. Direct and unmediated comportment is in fact the condition of possibility of all representation. Comportment, according to Heidegger (1927a), "makes possible every intentional relation to beings" and "precedes every possible mode of activity in general," prior to (explicit) beliefs, prior to (explicit) knowledge, and prior to (explicit) conceptual thinking (Heidegger, 1927a). That is, comportment is primary. The mistake of the traditional approaches lies in the fact that they treat (explicit) knowledge and its correlates as primary instead, and thus they turn the priority upside-down; in so doing, "every act of directing oneself toward something receives [wrongly] the characteristics of knowing" (Heidegger, 1927b). This is in essence the problem of classical cognitivism, including the difficulties with nature of representation, intentionality, consciousness, and so on. On the other hand, in the real world, agents "fix their beliefs not only in their heads but in their worlds, as they attune themselves differently to different parts of the world as a result of their experience" (Sanders, 1996).

Heidegger's philosophy eschewed the traditional internal/external dichotomy (Bechtel, 1988; Pollack, 1989) and, in its stead, posited a primordial way of interaction ("comportment"), direct and unmediated, as a foundation upon which cognitive processes (including high-level conceptual thinking) can exist. Being-inthe-world thus serves as a focal point of a different way of thinking about cognition. However, this is not at all to deny the existence of representation. To the contrary, high-level conceptual thinking involving (explicit) representation (as studied extensively in cognitive science and AI) does occur, but it is not as prevalent and as basic as readily assumed by classical cognitivism. It occurs only in more unusual circumstances.

It is useful to point out that the comportment that we are talking about is not exactly the same as "embodiment" that has been advocated as the key to understanding human cognition. Lakoff and Johnson (1980; Johnson, 1987) have been putting forth a view that cognition is largely determined by the bodily basis of cognitive agents: the bodily schemata get abstracted and mapped onto all other domains and all other cognitive processes. Interpreted based on that position, an object should not be understood and represented in terms of its shape, color, or any other static features, but should be approached in terms of what an agent can do with it (Glenberg, 1997). However, although these ideas are on the right track, they leave open too many possibilities. For example, there are too many different uses we can make of a cup: we can drink from it, we can use it to store coffee, tea, water, powder, coins, paper clips, or business cards, we can hold it in our hands, we can put it on top of our heads, we can stand on it, or we can play tricks with it. Its uses are unlimited. How can an agent structure its understanding around so many different possibilities? The key here, I believe, is what an agent usually, routinely, reflexively do with an object in its everyday life (amidst the "contexture of functionality"), i.e. the common and routine dealing with an object, which is what we call comportment (of an agent) toward an object (i.e. being-with-things; Heidegger, 1927a). Such comportment (or everyday routine dealing) with objects is the basis of how an agent approaches objects.

It is also important to note that in our view of comportment, there is no elaborate structure or network that explicitly encodes outcomes, enabling conditions, and other related information (as in Bickhard, 1998, which presents an alternative view). In some versions of situated cognition or interactivism (such as Bickhard, 1998; Cherian & Troxell, 1995), certain elaborate networks of pointers or other mechanisms are devised that relate different internal states in terms of their interrelations in action selection. Although they started out in the right direction, such schemes unfortunately fell right back into the representationalist trap and failed to capture the direct and unmediated nature of comportment. (For possible computational processes that are direct and unmediated, see the next section.)

Direct and unmediated comportment has been variously referred to as reactive routines, reactive skills, routine activities, everyday activities, ongoing interaction, everyday coping, and so on, in the work of, for example, Agre (1988), Chapman (1991), and Lave (1988). They have also been classified as subconceptual "representation" (in the broadest sense of the term "representation"; Smolensky, 1988; Sun, 1994). Brooks (1991), Maes and Brooks (1990), Agre (1988), and Chapman (1991) attempted to implement comportment in a variety of concrete ways, albeit without learning or development (more on learning later).

Conceptual processes and representation

Evidently, comportment is intentional only in the sense that it directs an agent to objects in the world as part of a "natural" structure (Merleau-Ponty, 1963). Such intentionality of an agent is not qualitatively different from that of a tennis ball or a car (see the discussion of both earlier). This kind of "pre-representational" (i.e. implicit) pattern of interaction with the world serves as the foundation of how an agent relates to its environment in its everyday activities and, more importantly, also serves as the foundation of more complex forms of intentionality. Explicit representation can only be formed on top of primordial comportment; explicit representation is secondary, or derivative, and its intentional content is derived from direct comportment with the world.

As argued in Sun (1997) and Sun et al. (in press), there is ample psychological evidence pointing to "bottom-up" learning that goes from comportment to conceptual, symbolic representation. Several lines of research demonstrate that agents can learn complex skilled activities without first obtaining (a large amount of) explicit knowledge (e.g. Berry & Broadbent, 1988; Karmiloff-Smith, 1986; Reber, 1989; Schacter, 1987; Stanley et al., 1989; Willingham et al., 1989). In research on implicit learning, Berry and Broadbent (1988), Willingham et al. (1989), and Reber (1989) demonstrate a dissociation between explicitly represented knowledge and performance in a variety of tasks, including dynamic control tasks (Berry & Broadbent, 1988), artificial grammar learning tasks (Reber, 1989), and serial reaction time tasks (Willingham et al., 1989). Berry and Broadbent (1988) argue that agents can learn to perform a task without being provided a priori explicit knowledge and without being able to explicitly verbalize the rules they used to perform the task. This indicates that skills are not necessarily accompanied by explicit knowledge. Willingham et al. (1989) similarly demonstrate that performance is not always preceded by explicit knowledge in human skill learning, and show that they are not necessarily correlated either. There are indications that explicit knowledge may arise from skilled activities in some circumstances (Stanley et al., 1989). Using a dynamic control task, Stanley et al. (1989) found that the development of explicit knowledge paralleled but lagged behind the development of skilled performance.

Similar claims concerning the development of performance prior to the development of explicit knowledge have surfaced in a number of other research areas and provided additional support for the bottom-up process. Implicit memory research (e.g. Schacter, 1987) demonstrates a dissociation between explicit and implicit memories in that an agent's performance can improve by virtue of implicit "retrieval" from memory and the agent can be unaware of the process. Instrumental conditioning is typically non-verbal and involves the formation of action sequences without explicit knowledge. It may be applied to simple organisms as well as humans (Gluck & Bower, 1988). In developmental psychology, Karmiloff-Smith (1986) proposes the idea of "representational redescription." During development, low-level implicit knowledge is transformed into explicit representation and thereby made more accessible. This process is bottom-up.

Note that the generation of low-level comportment during ontogenesis is determined by, at least, the following two important factors: (1) structures in the external world as perceived by the agent, which in turn depends on current structures in the agent and therefore on the (ontogenetic and phylogenetic) history of the agent/world interaction, and (2) innate "biases," or built-in constraints and predispositions, which also depend on the (ontogenetic and phylogenetic) history of the agent/world interaction. In turn, the generation of high-level structures (i.e. conceptual representation, with symbols) is, to a significant extent, determined by low-level structures, as well as sociocultural influences (especially signs/symbols existing and employed in a given culture) [10].

On this view, high-level conceptual, symbolic representation is rooted, or grounded, in low-level behavior (comportment) from which it obtains its meanings and for which it provides support and explanations. The rootedness/groundedness is guaranteed by the way high-level representation is produced: it is, in the main, extracted out of low-level behavioral structures. Even culturally transmitted symbols have to be linked up, within the mind of an individual agent, with low-level processes in order to be effective.

It is worth noting that conceptual, symbolic representation so formed is in general formed in a functionally relevant way, in relations to everyday activities of agents. In other words, in general, it must bear certain existential and/or biological significance to agents and be in the service of agents' activities. The world is of such a high (or even an infinitely high) dimensionality, and thus there can be no totally objective way of perceiving/conceiving it (due to the complexity), except in relation to what an agent has to do with it in everyday activities. Learning symbolic representation on the basis of comportment provides agents with a viable way of basing their conceptual representation on their everyday activities in a functionally relevant way [11].

The existence of explicit representation, or at least its importance, has been denied, or downplayed, by many advocates of (strong forms of) situated cognition and interactivism (e.g. Bickhard, 1993; Brooks, 1991; Port & van Gelder, 1995; Suchman, 1987). The existence of explicit representation (but not its paramount role) has in fact been argued for by a number of researchers, persuasively, I believe. See, for example, Markman and Dietrich (1998) and Smith et al. (1992). Here I take the more "eclectic" position that acknowledges that representation is important while maintaining that it is mediated by direct comportment.

A dual process theory

This analysis boils down to the dual process theory (i.e. the dual-level hypothesis; Sun, 1994). In one of my previous books (entitled Integrating rules and connectionism for robust commonsense reasoning; Sun, 1994), 1 put forth the following hypothesis (in which the word "knowledge" should be broadly interpreted):

It is assumed in this work that cognitive processes are carried out in two distinct levels with qualitatively different processing mechanisms. Each level encodes a fairly complete set of knowledge for its processing, and the coverage of the two sets of knowledge encoded by the two levels overlaps substantially.

This idea is closely related to some well-known dichotomies in cognitive science: the dichotomy of symbolic versus subsymbolic processing (Rumelhart & McClelland, 1986), the dichotomy of conceptual versus subconceptual processing (Smolensky, 1988), the dichotomy of explicit versus implicit learning (Berry & Broadbent, 1988; Reber, 1989), the dichotomy of controlled versus automatic processing (Schneider & Oliver, 1991; Shiffrin & Schneider, 1977), and the dichotomy of declarative and procedural knowledge (Anderson, 1983). However, different from some of these dichotomies, I went further in positing separate and simultaneous existence of multiple levels (i.e. separate processors), each of which embodies one side of a dichotomy (cf. Grossberg, 1987). Therefore, in this work, the two sides of a dichotomy are not simply two ends of a spectrum, or two levels of analysis of the same underlying system. But they are two separate, although closely connected, systems.

In the current context, the two distinct levels are termed, respectively, the conceptual level and the subconceptual level, following the usage by Smolensky (1988). The two levels encode similar and comparable content (or "knowledge" in a broad sense). But they encode their content in different ways. One works in a comportment-like way while the other in an explicit, symbolic, and conceptual way. They also utilize different processing mechanisms. Thus, they can have qualitatively different flavors, although they can function together. The reason for having the two levels, or any other similar combination of components, is that these different levels can potentially work together synergistically, supplementing and complementing each other in a variety of different ways (as demonstrated in Sun, 1994, 1997; Sun & Peterson, 1998a, b) [12]. I have argued for the two-level hypothesis extensively before, based on a variety of evidence (see e.g. Sun, 1995, 1997). So I will not repeat the arguments here.

Computational analysis of everyday activities

Let me briefly sketch a picture of the mechanistic underpinning of this theory, that is, an architecture for a model of the mind [13]. I want to put together here some basic ingredients. First of all, the new approach should start with only minimum built-in initial structures in an agent. Some of these initial structures have to do with "pre-wired" reflexes, or predisposition for developing such reflexes, that is, genetic and biological pre-endowment. Some others have to do with learning capabilities, since most of the structures in cognition will have to be constructed in a gradual, incremental fashion, during the course of individual ontogenesis (so one might view this as a constructivist approach, although different from the Piagetian approach; cf. Infielder & Piaget, 1958). The development of structures is based on interaction with the world (i.e. based on being-in-the-world, including both the physical world and the social/cultural world). The interaction prompts the formation of various low-level structures in behavior, which in turn lead to the emergence of high-level conceptual representation.

In developing this approach, connectionist models are utilized, in a properly generalized form, as the basic unifying medium of implementation, and as a guiding metaphor for constructing hypotheses, models, and theories (Sun, 1994, 1995; Sun & Bookman, 1994), because of the many appealing properties of such models (Bechtel & Abrahamsen, 1991; Clark, 1993; Smolensky, 1988) and because of the fact that it encompasses both symbolic approaches and dynamical system approaches (Port & van Gelder, 1995).

Computational processes of comportment

Now we are ready to probe these ideas further. First, we need to gain a better understanding of comportment, beyond mere philosophical speculation. In this subsection, we shall examine the computational processes of comportment. We shall also look into the development of comportment in the "ontogenesis" of an individual agent, which is the most important means by which comportment is acquired (although some innate structures might be formed evolutionarily, a priori, as mentioned before).

Generally speaking, while performing everyday activities (especially in direct comportment), the agent is under time pressure: often, a mundane action "decision" has to be made in a fraction of a second; it cannot involve much of "information processing" and falls outside of Allen Newell's "rational band" (i.e. cognitive processes that take minutes or hours to complete, which is what cognitive science and AI traditionally deal with). The agent is also severely limited in other resources, such as memory, so that memorizing and analyzing all the previous experience (in detail) are impossible (although some form of episodic memory obviously exists). The perceptual ability of the agent is also limited in that only local information is available. Goals may not be explicit and a priori to an agent either. They may be implied in reinforcements/payoffs received, and they are pursued by an agent as a side-effect of trying to maximize reinforcements/payoffs.

Learning of comportment is an experiential, trial-and-error process; the agent develops its competence, tentatively, on an ongoing basis (because it cannot wait until the end of its experience before making a decision and starting to learn) [14]. In general, as demonstrated by the models of Nosofsky et al. (1994) and Medin et al. (1987), human learning is mostly gradual, ongoing, and concurrent (online), which is especially true in learning comportment. The characteristics of the world, from the viewpoint of the learning agent, need not be stationary. It can be nonstationary ("drifting") in several ways: (1) the world can change over time; thus, the revision of comportment structures learned by an agent may be necessary; (2) even when the world per se is stationary, it may still seem evolving to an agent learning to cope with the world, because different regions of the world may exhibit different characteristics and thus revisions over time may be required (Widmer & Kubat, 1996). In general, there is no preselected set of instances that provide a fixed view of the world; (3) once a structure is revised, the agent has to view whatever it experienced before in new ways (because the current comportment structures serve as a "filter" through which the agent sees the world), and thus the experience may seem different and the world nonstationary; and (4) there is a lack of a clear and steady criterion for learning comportment. Reinforcement/payoffs may be received sporadically, and it is up to the agent to decide what to make of them. The agent has to assign credits/blames on the basis of what is already known, which is constantly changing. Since the learning criterion is a moving target, the learning process becomes nonstationary.

There are some existing computational methods available to accomplish simple forms of such learning. Chief among them is the temporal difference method (Sutton, 1988), a type of reinforcement learning that learns through exploiting the difference in evaluating actions in successive steps and thus handling sequences in an incremental fashion. Another approach, genetic algorithm (Holland et al., 1986), may also be used to tackle this kind of task. As an example, to implement the first approach using neural nets, as discussed in Sun (1997), we can use a four-layered network in which the first three layers form a backpropagation network (feedforward or recurrent; Rumelhart & McClelland, 1986) for computing output action values (i.e. Q-values; Watkins, 1989) and the fourth layer performs stochastic decision making (Watkins, 1989). The network can be internally subsymbolic, involving distributed features as developed automatically through the backpropagation algorithm. The output of the third layer indicates the value of each action (represented by an individual node). The value is an evaluation of the "quality" of an action in a given input state. To acquire these values, we can use the standard Q-learning algorithm (a temporal difference reinforcement learning algorithm as mentioned earlier). It basically compares the values of successive actions and adjusts an evaluation function on that basis, without explicitly involving and representing goals, states, and outcomes. (For details of Q-learning, see Watkins, 1989, as well as more recent treatments such as Kaelbling et al., 1996.) The afore-specified model is in fact one part of our overall model (named CLARION), which is specifically concerned with comportment (Sun, 1997; Sun & Peterson, 1998a,b). Clearly, in this implementation of comportment, there is no elaborate internal representation, and no pointer between different entities representing mental states (cf.Bickhard, 19980, as has been discussed eaarlier.

Computational processes of conceptual processing

Let us now discuss computational processes of high-level conceptual processing, with (explicit) representation, during agents' everyday activities.

Let us see how explicit representation is acquired. We have discussed the idea of "bottom-up" learning. But how do we accomplish "bottom-up" learning computationally? Admittedly, there are many symbolic rule learning algorithms out there for learning explicit rules (as a form of high-level explicit representation). However, the afore-identified characteristics of everyday activities render most existing rule learning algorithms inapplicable, because they require either preconstructed exemplar sets (Michalski, 1983; Quinlan, 1986, 1990), incrementally given consistent instances (Fisher, 1987; Mitchell, 1982; Utgoff, 1989), or complex manipulations of learned symbolic structures when inconsistency is discovered (which is typically more complex than the limited time an agent may have in reactive activities; Hirsh, 1994). "Drifting" as analyzed before is clearly more than noise and inconsistency considered by some learning algorithms, because it involves changes over time and may lead to radical changes in learned knowledge. Above all, most of the rule learning algorithms do not handle the learning of sequences, which is an essential form of agents' everyday activities and necessarily involves temporal credit assignment.

Therefore, algorithms have to be developed specifically for the purpose of modeling the acquisition of explicit representation in agents (Sun & Peterson, 1998a,b). The algorithms should be bottom-up; that is, they utilize whatever is learned in the lower level (the part of the computational model that implements comportment, described in the previous subsection) and construct symbolic representation at the higher level (the part that implements conceptual processing in agents). The basic idea for such bottom-up learning is as follows: if some action decided by the bottom level of an agent is successful (being successful could mean a number of different things; see Sun & Peterson, 1998a,b, regarding details), then the agent extracts an explicit rule that corresponds to the action selected by the bottom level and adds the rule to the top level. Then, in subsequent interaction with the world, the agent verifies the extracted rule by considering the outcome of applying the rule: if the outcome is not successful, then the rule should be made more specific and exclusive of the current case; if the outcome is successful, the agent may try to generalize the rule to make it more universal. (Specialization and generalization are done based on actually encountered situations, and thus search is minimal.) To measure the success or failure of a step in order to learn rules, some statistical criteria (based on information gain) have been developed and tested in Sun and Peterson (1998a,b).

We use a localist connectionist model for representing these rules (where "localist" means that each concept is represented separately by an individual node in a network). Basically, we connect the nodes representing conditions of a rule to the node representing the conclusion (through standard sigmoid activation functions). That is, we translate the structure of a set of rules into the structure of a network (see e.g. Sun, 1994, 1995, for details of localist encoding). The rule learning algorithm (as described above) is implemented on top of the localist network (Sun & Peterson, 1998a,b). The use of the learned rules is through a number of inference algorithms (such as backward chaining and forward chaining), which are also implemented on top of the network (through controlling activation propagation; Sun, 1997).

Besides bottom-up influences, there are also "top-down" influences (in learning and otherwise). In learning, such influences have been amply studied and modeled by top-down learning models such as ACT (Anderson, 1983; Anderson & Lebiere,1998). In our framework, when a rule is given, the agent can represent it in the top-level rule network and connect it to existing representation [15]. We can go one step further by performing rule assimilation, through which rules given externally, besides being represented in the top level, are assimilated into the bottom level and thus become more effective (see Anderson, 1983; Dreyfus & Dreyfus, 1987). We can train the bottom level according to rules given in the top level, using supervised learning (e.g. backpropagation). Aside from learning, top-down influences are apparent in performance as well. For example, Rips (1989) showed that even categorization does not rely on just similarity but also reasoning from rules. The computational processes of top-down influences can be captured through combining the outcomes of the two levels in making decisions (Sun, 1997).

The afore-described ideas concerning computational processes of both comportment (subconceptual processing) and conceptual processing have been implemented in a (computational) cognitive model named CLARION. The model has been described extensively in a series of papers, including Sun (1997), Sun and Peterson (1998a,b), and Sun et aL (1998a,b, in press). Essentially, as described above, it is made up of two levels, the top level conceptual and the bottom level subconceptual (comportment-oriented) [16]. The two levels interact in action selection, through combining action recommendations from the two levels, respectively, and they cooperate in learning through the afore-described bottom-up and top-down processes. The model grounds symbols and symbolic representation through bottom-up learning and through comportment with the world.

Concept formation

While learning rules, CLARION forms concepts (Sun & Peterson, 1998b). Although there are available distributed features in the bottom level for specifying rule conditions, a separate node is instead set up in the top level to represent the condition of a rule that connects to the distributed features. So along with induced rules, localist (i.e. symbolic) representation of concepts is formed. Each localist node is linked to its corresponding distributed features in the bottom level, from which it was abstracted, through bottom-up activation (using a standard sigmoid activation function with uniform weights).

This kind of concept representation is basically a prototype model (Rosch, 1978; Smith & Medin, 1981). Localist nodes serve as identification and encoding of features, in a bottom-up direction. They also serve to trigger relevant distributed features, in a top-down direction, once a concept is brought into attention [17].

Moreover, concepts formed in this way in CLARION are context-dependent and task-oriented, because they are always formed with regard to the tasks at hand while exploiting environmental regularities (Heidegger, 1927a; Okrent, 1996). The representation of an agent, for the most part, need not be determined a priori. Being emphasized in CLARION is the functional role of concepts and the importance of function in forming concepts. A concept is formed as part of a rule, which is learned to accomplish a task in a certain environment. Therefore, acquired concepts are functional. The task context and the experience help an agent to determine which features in the environment need to be emphasized, what stimuli should be grouped together, and thus what constitutes a separate category (that is, a concept). (See Sun & Peterson, 1998b, for a more detailed analysis of concepts formed by agents during learning of specific tasks.) This may explain why most human concepts are (more or less) concerned with existentially and biologically significant aspects of the world; they are not just "objective" classifications (Lakoff & Johnson, 1980; Lave, 1988).

Such concept formation and representation may have interesting implications for the frame problem in AI. The frame problem refers to the difficulty in keeping track of many propositions concerning a situation that may be in a constant flux. Any change in the situation may lead to the validation of many propositions and the invalidation of many others. This may lead to very high computational complexity when we need to reason about the situation. The idea of "tracking" was envisaged in the purely symbolic representational framework of traditional AI, which makes such a process necessary. However, concepts being grounded, context-dependent, and task-oriented (such as in CLARION) may alleviate the need for such "tracking." In this alternative approach, each situation is reacted to, first and foremost, by low-level comportment and then, through low-level comportment, it triggers proper concepts and propositions at the higher level, which thus produces inferences that are tailored to the situation (Sun, 1994). Purely logical reasoning concerning each and every proposition possible (as envisaged by traditional approaches) is computationally excessive and is rendered unnecessary by this approach (Sun, 1994, 1995). (Certainly, we need much more work along this direction to further elucidate the potential of this alternative view.)

Representation and intentionality: an assessment

Now we can revisit the issue of representation. I was critical of the traditionally dominant and currently lingering position on representation in cognitive science. However, rejecting representationalism does not necessarily mean rejecting (weak notions of) representation and symbols.

Let us examine CLARION. In this model, the top level indeed consists of representation in the sense implied by representationalism; this is because encoding used there is punctate (with each item being in an isolatable node) and also elaborate (with represented items forming a rather complete model of a domain, if CLARION is given enough time to learn). Moreover, the representation is symbolic, in the sense that a concept is assigned to an arbitrary node, without any intrinsic connection between a node and what it represents. Syntactic structures (concatenative compositional structures; Fodor, 1975) can be built on the basis of such representation. Syntactically sensitive symbolic processing can be performed on them. This level of the model is thus representational. However, the bottom level of CLARION is different. A connectionist model (namely, a backpropagation network; Rumelhart & McClelland, 1986) is used; thus distributed feature encoding is involved. In such a scheme, there is no symbol, that is, no lexical item that can be assigned arbitrary meanings. Moreover, it is not a priori determined (beside certain presumably biologically built-in constraints). The encoding does not exist before an agent learns to interact with the world and thus develops; it is intrinsically tied to the experience of interaction between an agent and its environment. There is no syntactic structure in that level (in the sense of concatenative compositional structures; Fodor, 1975). Thus this level of CLARION is non-representational. Putting the two levels together, CLARION incorporates both representation and non-representation. However, the model does not simply juxtapose the two qualitatively different components, but combines them in an integral framework, so that the two parts of the model can interact and cooperate [18].

What is the implication of the above discussion for the question of intentionality? Let us compare how the meanings of the contents of the bottom level and the top level are determined. To understand this issue, the notions of intrinsic intentionality and derived intentionality (Searle, 1980) are pertinent. According to Heidegger (1927b), representation presupposes the more basic comportment with the world. Comportment carries with it a direct and unmediated relation to, and reference of, things in the world; being-with-things is a fundamental part of being-in-the-world, a bridge to the existential context of an agent. Therefore, it provides an intrinsic intentionality (meanings), or in other words, a connection (to things in the world) that is intrinsic to an agent, given a particular existential context of the agent (and its biological pre-endowment). In addition to intrinsic intentionality, there is derived intentionality, which is obtained through derivative means. In CLARION, intentionality can be categorized into these two kinds precisely: the bottom-level processes that capture direct comportment with the world and the top-level processes that are the result of extracting rules and concepts from the bottom-level processes. The bottom-level processes acquire their internal encoding through learning from the experience of direct interaction with the world, and thus the meanings of the encoding lie in the intrinsicness of the weights and the wiring, which are determined by the process of unmediated interaction (through Q-learning and backpropagation in the model) [19]. The top-level processes result from extraction, i.e. derivation. Thus the meanings or intentionality of the representation can only be traced to the derivation process. Through derivation/extraction, as well as through the ongoing connection to the bottom level (both top-down and bottomup connections), symbols at the top level are "grounded" in the bottom level and, through the bottom level, in the comportment with the world. I want to emphasize that not only symbols themselves are derived from the bottom-level processes but the meanings of these symbols are thus also derived from these processes.

Moreover, although Heidegger recognized the ontological precedence of intrinsic intentionality, it is also important to further recognize, as in our model, that intrinsic intentionality is not only ontologically prior to derived intentionality, but also developmentally (ontogenetically) prior to it, in individual agents. As demonstrated by Inhelder and Piaget (1958), a child learns concepts and schemas only when the child learns to interact with objects in the world, and increasingly and more complex concepts are developed through increasingly more complex interaction with objects. As suggested by Karmiloff-Smith (1992), the increasing mastery of concepts are accomplished, in part, through a "representational redescription" process: first, a child acquires an embodied performance ability, then through representational redescription, i.e. extracting explicit representation, the child learns explicit concepts and thereby further improves performance. CLARION, as described earlier, roughly captures this developmental process (in a qualitative way; Sun, 1997).

Further discussions

Comparisons

We can contrast the afore-outlined approach with some traditional thinking on cognition. One major difference we see is that traditional thinking tends to overlook various external factors in cognition. David Hume (Hume, 1938) believed that cognition can only be understood on the basis of sense data that an individual agent receives, based on which an agent forms associations that relate them, in order to make sense of them. In William James's The principles of psychology (James, 1890), in a total of 28 chapters covering a wide-ranging set of topics, cognition is construed as merely an internal process that works on data provided by external sources. In Readings in cognitive science, edited by Collins and Smith (1988), a major collection of significant early work in cognitive science, the field of cognitive science was defined to be "the interdisciplinary study of the acquisition and use of knowledge." Despite the fact that knowledge is the result of agents' interaction with the world (individually and/or collectively), there is no treatment of such dynamic ongoing interaction in the book.

In contemporary cognitive science and AI, although ideas similar to some of those outlined in the present framework have started to seep into various segments of different research communities (see e.g. Agre, 1995; Damasio, 1994; Hutchins, 1995; Russell & Norvig, 1995; Sun, 1994), cognitive science/Al, as a whole, has not been particularly hospitable to these new ideas. See, for example, Vera and Simon (1993) and Hayes et al. (1994) for a glimpse of the opposing views.

Another contrast is with regard to the duality in cognition (the dual processes), which has long been speculated upon. Although the notion of the conscious versus the subconscious has captivated pop culture ever since Freud (1937; Kitcher, 1992), in mainstream academic psychology (especially cognitive psychology) and in mainstream academic philosophy, it is not quite readily accepted (although there are some notable exceptions). The distinction of the conceptual versus the subconceptual was proposed, in the context of analyzing connectionist models (Smolensky, 1988), as a sort of substitute for the distinction of the conscious versus the subconscious (to avoid the controversies surrounding the latter). The distinction of the conceptual versus the subconceptual has not been very popular either. In contrast, we take such dichotomies seriously and adopt them as the basis of our approach.

One additional note should be made regarding the relation between our dichotomy and the dichotomy of declarative versus procedural knowledge (Anderson, 1983), which came closest to our dichotomy. The two dichotomies are very similar, except that Anderson's model did not account well for bottom-up learning, because it is based (mostly) on top-down learning (to capture various instructed learning situations), and thus it did not account for the derivation and the grounding of symbols and representation.

The approach outlined here is consistent with the situated cognition view (interactivism), in the sense that coping with the world means acting in an environmentally driven fashion and dealing with moment-to-moment contingencies. Our approach reflects such a view through a focus on reacting to the current state of the world. Also in line with the situated cognition view and interactivism, learning in our approach is tied closely to specific situations as experienced (to reflect and exploit environmental contingencies and regularities). But there are some obvious differences. The situated cognition view often claims that there should not be any elaborate model of the world or elaborate representation. However, instead of being completely antithetical to the representationalist view and hastily avoiding any representation or model, we take a more inclusive approach: we show that explicit representation can be constructed on the basis of situated learning by situated agents through a bottom-up process, thus unifying the two contradictory views.

Concluding remarks

This article shows how representation and representational content emerge in the interaction between an agent and its environment. It hypothesizes the process that goes from unmediated comportment with the world to mediated (symbolic and conceptual) representation and the concomitant conceptual processing.

Our framework outlined above reconciles representationalism and situated cognition interactivism. It does so through explicating the crucial role played by direct and unmediated comportment. Comportment bridges the gap between the world and the internal (explicit) representation in an agent. It makes (explicit) representation possible by giving rise to it through experience.

The key of our work lies in recognizing the fact that experiences (everyday activities) come first and representation comes later (which has been recognized by many) and, based on that, constructing computational models that demonstrate the feasibility of this view (which is novel). Instead of Descartes' motto "I think, therefore I am" (or the revisionist version "I feel, therefore I am"; Damasio, 1994), we now argue that (computationally) it can be and should be "I am, therefore I think." I believe that this reversal sets the priority right for computational cognitive modeling, as well as for cognitive science in general.