PERCEPTUAL SYMBOL SYSTEMS

Keywords

analogue processing, categories, concepts, frames, imagery, images, knowledge, perception, representation, sensory-motor representations, simulation, symbol grounding, symbol systems

Abstract

Prior to the twentieth century, theories of knowledge were inherently perceptual. Since then, developments in logic, statistics, and programming languages have inspired amodal theories that rest on principles fundamentally different from those underlying perception. In addition, perceptual approaches have become widely viewed as untenable, because they are assumed to implement recording systems, not conceptual systems. A perceptual theory of knowledge is developed here in the contexts of current cognitive science and neuroscience. During perceptual experience, association areas in the brain capture bottom-up patterns of activation in sensory-motor areas. Later, in a top-down manner, association areas partially reactivate sensory-motor areas to implement perceptual symbols. The storage and reactivation of perceptual symbols operates at the level of perceptual components--not at the level of holistic perceptual experiences. Through the use of selective attention, schematic representations of perceptual components are extracted from experience and stored in memory (e.g., individual memories of green, purr, hot). As memories of the same component become organized around a common frame, they implement a simulator that produces limitless simulations of the component (e.g., simulations of purr). Not only do such simulators develop for aspects of sensory experience, they also develop for aspects of proprioception (e.g., lift, run) and for introspection (e.g., compare, memory, happy, hungry). Once established, these simulators implement a basic conceptual system that represents types, supports categorization, and produces categorical inferences. These simulators further support productivity, propositions, and abstract concepts, thereby implementing a fully functional conceptual system. Productivity results from integrating simulators combinatorially and recursively to produce complex simulations. Propositions result from binding simulators to perceived individuals to represent type-token relations. Abstract concepts are grounded in complex simulations of combined physical and introspective events. Thus, a perceptual theory of knowledge can implement a fully functional conceptual system while avoiding what it is becoming increasingly apparent would be problems for amodal symbol systems. Implications for cognition, neuroscience, evolution, development, and artificial intelligence are explored.

The habit of abstract pursuits makes learned men much inferior to the average in the power of visualization, and much more exclusively occupied with words in their 'thinking'.

Bertrand Russell (1919b)

For the last several decades, the fields of cognition and perception have diverged. Researchers in these two areas know ever less about each other's work, and their discoveries have had diminishing influence on each other. In many universities, researchers in these two areas are in different programs, and sometimes in different departments, buildings, and university divisions. One might conclude from this lack of contact that perception and cognition reflect independent or modular systems in the brain. Perceptual systems pick up information from the environment and pass it on to separate systems that support the various cognitive functions, such as language, memory, and thought. I will argue that this view is fundamentally wrong. Instead, cognition is inherently perceptual, sharing systems with perception at both the cognitive and the neural levels. I will further suggest that the divergence between cognition and perception reflects the widespread assumption that cognitive representations are inherently nonperceptual, or what I will call amodal.

1.1. Grounding Cognition in Perception

In contrast to modern views, it is relatively straightforward to imagine how cognition could be inherently perceptual. As Figure 1 illustrates, this view begins by assuming that perceptual states arise in sensory-motor systems. As discussed in more detail later (2.1), a perceptual state can contain two components: an unconscious neural representation of physical input, and an optional conscious experience. Once a perceptual state arises, a subset of it is extracted via selective attention and stored permanently in long-term memory. On later retrievals, this perceptual memory can function symbolically, standing for referents in the world, and entering into symbol manipulation. As collections of perceptual symbols develop, they constitute the representations that underlie cognition.

Perceptual symbols are modal and analogical. They are modal because they are represented in the same systems as the perceptual states that produced them. The neural systems that represent color in perception, for example, also represent the colors of objects in perceptual symbols, at least to a significant extent. On this view, a common representational system underlies perception and cognition, not independent systems. Because perceptual symbols are modal, they are also analogical. The structure of a perceptual symbol corresponds, at least somewhat, to the perceptual state that produced it.¹

Given how reasonable this perceptually based view of cognition might seem, why has it not enjoyed widespread acceptance? Why is it not in serious contention as a theory of representation? Actually, this view dominated theories of mind for most of recorded history. For over 2,000 years, theorists viewed higher cognition as inherently perceptual. Since Aristotle (4th century BC/1961) and Epicurus (4th century BC/1994), theorists saw the representations that underlie cognition as imagistic. British empiricists such as Locke (1690/1959), Berkeley (1710/1982), and Hume (1739/1978) certainly viewed cognition in this manner. Images likewise played a central role in the theories of later nativists such as Kant (1787/1965) and Reid (1764/1970, 1785/1969). Even recent philosophers such as Russell (1919b) and Price (1953) have incorporated images centrally into their theories. Until the early twentieth century, nearly all theorists assumed that knowledge had a strong perceptual character.

After being widely accepted for two millennia, this view withered with mentalism in the early twentieth century. At that time, behaviorists and ordinary language philosophers successfully banished mental states from consideration in much of the scientific community, arguing that they were unscientific and led to confused views of human nature (e.g., Ryle, 1949; Watson, 1913; Wittgenstein, 1953). Because perceptual theories of mind had dominated mentalism to that point, attacks on mentalism often included a critique of images. The goal of these attacks was not to exclude images from mentalism, however, but to eliminate mentalism altogether. As a result, image-based theories of cognition disappeared with theories of cognition.

1.2. Amodal Symbol Systems

Following the cognitive revolution in the mid-twentieth century, theorists developed radically new approaches to representation. In contrast to pre-twentieth century thinking, modern cognitive scientists began working with representational schemes that were inherently nonperceptual. To a large extent, this shift reflected major developments outside cognitive science in logic, statistics, and computer science. Formalisms such as predicate calculus, probability theory, and programming languages became widely known and inspired technical developments everywhere. In cognitive science, they inspired many new representational languages, most of which are still in widespread use today (e.g., feature lists, frames, schemata, semantic nets, procedural semantics, production systems, connectionism).

These new representational schemes differed from earlier ones in their relation to perception. Whereas earlier schemes assumed that cognitive representations utilize perceptual representations (Figure 1), the newer schemes assumed that cognitive and perceptual representations constitute separate systems that work according to different principles. Figure 2 illustrates this assumption. As in the framework for perceptual symbol systems in Figure 1, perceptual states arise in sensory-motor systems. However, the next step differs critically. Rather than extracting a subset of a perceptual state and storing it for later use as a symbol, an amodal symbol system transduces a subset of a perceptual state into a completely new representation language that is inherently non-perceptual.

As amodal symbols become transduced from perceptual states, they enter into larger representational structures, such as feature lists, frames, schemata, semantic networks, and production systems. These structures in turn constitute a fully functional symbolic system with a combinatorial syntax and semantics, which supports all of the higher cognitive functions, including memory, knowledge, language, and thought. For general treatments of this approach, see Dennett (1969), Newell and Simon (1972), Fodor (1975), Pylyshyn (1984), and Haugeland (1985). For reviews of specific theories in psychology, see E. Smith and Medin (1981), Rumelhart and Norman (1988), and Barsalou and Hale (1993).

It is essential to see that the symbols in these systems are amodal and arbitrary. They are amodal because their internal structures bear no correspondence to the perceptual states that produced them. The amodal symbols that represent the colors of objects in their absence reside in a different neural system from the representations of these colors during perception itself. In addition, these two systems use different representational schemes and operate according to different principles.

Because the symbols in these symbol systems are amodal, they are linked arbitrarily to the perceptual states that produce them. Analogous to how words typically have arbitrary relations to entities in the world, amodal symbols have arbitrary relations to perceptual states. Just as the word "chair" has no systematic similarity to physical chairs, the amodal symbol for chair has no systematic similarity to perceived chairs. As a consequence, similarities between amodal symbols are not related systematically to similarities between their perceptual states, which is again analogous to how similarities between words are not related systematically to similarities between their referents. Just as the words "blue" and "green" are not necessarily more similar than the words "blue" and "red," the amodal symbols for blue and green are not necessarily more similar than the amodal symbols for blue and red.²

Amodal symbols bear an important relation to words and language. Theorists typically use linguistic forms to represent amodal symbols. In feature lists, words represent features, as in:

CHAIR                               (1)
    seat
    back
    legs

Similarly in schemata, frames, and predicate calculus expressions, words represent relations, arguments, and values, as in:

Although theorists generally assume that words do not literally constitute the content of these representations, it is assumed that close amodal counterparts of words do. Although the word "horse" does not represent the value of Agent for EAT in (2), an amodal symbol that closely parallels this word does. Thus, symbolic thought is assumed to be analogous in many important ways to language. Just as language processing involves the sequential processing of words in a sentence, so conceptual processing is assumed to involve the sequential processing of amodal symbols in list-like or sentence-like structures (e.g., Fodor & Pylyshyn, 1988).

It is important to see that this emphasis on amodal and arbitrary symbols also exists in some, but not all, connectionist schemes for representing knowledge (e.g., McClelland, Rumelhart, & the PDP Research Group, 1986; Rumelhart, McClelland, & the PDP Research Group, 1986). Consider a feed-forward network with back propagation. The input units in the first layer constitute a simple perceptual system that codes the perceived features of presented entities. In contrast, the internal layer of hidden units is often interpreted as a simple conceptual system, with a pattern of activation providing the conceptual representation of an input pattern. Most importantly, the relation between a conceptual representation and its perceptual input is arbitrary for technical reasons. Prior to learning, the starting weights on the connections between the input units and the hidden units are set to small random values (if the values were all 0, the system couldn't learn). As a result, the conceptual representations that develop through learning are related arbitrarily to the perceptual states that activate them. With different starting weights, arbitrarily different conceptual states correspond to the same perceptual states. Even though connectionist schemes for representation differ in important ways from more traditional schemes, they often share the critical assumption that cognitive representations are amodal and arbitrary.

Connectionist representational schemes need not necessarily work this way. If the same associative network represents information in both perception and cognition, it grounds knowledge in perception and is not amodal (e.g., Pulvermüller, in press). As described later (2.2.1, 2.5), shared associative networks provide a natural way to view the representation of perceptual symbols.

1.2.1. Strengths. Amodal symbol systems have many powerful and important properties that any fully functional conceptual system must exhibit. These include the ability to represent types and tokens, to produce categorical inferences, to combine symbols productively, to represent propositions, and to represent abstract concepts. Amodal symbol systems have played the critical role of making these properties central to theories of human cognition, making it clear that any viable theory must account for them.

1.2.2. Problems. It has been less widely acknowledged that amodal symbol systems face many unresolved problems. First, there is little direct empirical evidence that amodal symbols exist. Using picture and word processing tasks, some researchers have explicitly tested the hypothesis that conceptual symbols are amodal (e.g., Snodgrass, 1984; Theios & Amhrein, 1989). However, a comprehensive review of this work concluded that conceptual symbols have a perceptual character (Glaser, 1992; also see Seifert, 1997). More recently, researchers have suggested that amodal vectors derived from linguistic context underlie semantic processing (Burgess & Lund, 1997; Landauer & Dumais, 1997). However, Glenberg et al. (1998b) provide strong evidence against these views, suggesting instead that affordances derived from sensory-motor simulations are essential to semantic processing.

Findings from neuroscience also challenge amodal symbols. Much research has established that categorical knowledge is grounded in sensory-motor regions of the brain (for reviews see Damasio, 1989; Gainotti, Silveri, Daniele, & Giustolisi, 1995; Pulvermüller, in press; also see 2.3). Damage to a particular sensory-motor region disrupts the conceptual processing of categories that use this region to perceive physical exemplars. For example, damage to the visual system disrupts the conceptual processing of categories whose exemplars are primarily processed visually, such as birds. These findings strongly suggest that categorical knowledge is not amodal.³

In general, the primary evidence for amodal symbols is indirect. Because amodal symbols can implement conceptual systems, they receive indirect support through their instrumental roles in these accounts. Notably, however, amodal symbols have not fared well in implementing all computational functions. In particular, they have encountered difficulty in representing spatio-temporal knowledge, because the computational systems that result are cumbersome, brittle, and intractable (e.g., Clark, 1997; Glasgow, 1993; McDermott, 1987; Winograd & Flores, 1987). Although amodal symbol systems do implement some computational functions elegantly and naturally, their inadequacies in implementing others is not encouraging.

Another shortcoming of amodal symbol systems is their failure to provide a satisfactory account of the transduction process that maps perceptual states into amodal symbols (Figure 2). The lack of an account for such a critical process should give one pause in adopting this general framework. If we can not explain how these symbols arise in the cognitive system, why should we be confident that they exist? Perhaps even more serious is the complete lack of cognitive and neural evidence that such a transduction process actually exists in the brain.

A related shortcoming is the symbol grounding problem (Harnad, 1987; 1990; Newton, 1996; Searle, 1980), which is the converse of the transduction problem. Just as we have no account of how perceptual symbols become mapped to amodal symbols during transduction, neither do we have an account of how amodal symbols become mapped back to perceptual states and entities in the world. Although amodal theories often stress the importance of symbol interpretation, they fail to provide compelling accounts of the interpretive scheme that guides reference. Without such an account, we should again have misgivings about the viability of this approach.⁴

A related problem concerns how an amodal system implements comprehension in the absence of physical referents. Imagine that amodal symbols are manipulated to envision a future event. If nothing in the perceived environment grounds these symbols, how does the system understand its reasoning? Because the processing of amodal symbols is usually assumed to be entirely syntactic (based on form and not meaning), how could such a system have any sense of what its computations are about? It is often argued that amodal symbols acquire meaning from associated symbols, but without ultimately grounding terminal symbols, the problem remains unsolved. Certainly, people have the experience of comprehension in such situations.

One solution is to postulate mediating perceptual representations (e.g., Harnad, 1987; Höffding, 1891; Neisser, 1967). According to this account, every amodal symbol is associated with corresponding perceptual states in long-term memory. For example, the amodal symbol for dog is associated with perceptual memories of dogs. During transduction, the perception of a dog activates these perceptual memories, which activate the amodal symbol for dog. During symbol grounding, the activation of the amodal symbol in turn activates associated perceptual memories, which ground comprehension. Problematically, though, perceptual memories are doing all of the work, and the amodal symbols are redundant. Why couldn't the system simply use its perceptual representations of dogs alone to represent dog, both during categorization and reasoning?

The obvious response from the amodal perspective is that amodal symbols perform additional work that these perceptual representations cannot perform. As we shall see, however, perceptual representations can play the critical symbolic functions that amodal symbols play in traditional systems, such that amodal symbols become redundant. If we have no direct evidence for amodal symbols, as noted earlier, then why postulate them?

Finally, amodal symbol systems are too powerful. They can explain any finding post hoc (Anderson, 1978), but often without providing much illumination. Besides being unfalsifiable, these systems often fail to make strong a priori predictions about cognitive phenomena, especially those of a perceptual nature. For example, amodal theories do not naturally predict distance and orientation effects in scanning and rotation (Finke, 1989; Kosslyn, 1980), although they can explain them post hoc. Such accounts are not particularly impressive, though, because they are unconstrained and offer little insight into the phenomena.

1.2.3. Theory evaluation. Much has been made about the ability of amodal theories to explain any imagery phenomenon (e.g., Anderson, 1978). However, this ability must be put into perspective. If perceptual theories predict these effects a priori, whereas amodal theories explain them post hoc, why should this be viewed as a tie? From the perspective of inferential statistics, Bayesian reasoning, and philosophy of science, post hoc accounts should be viewed with great caution. If a priori prediction is favored over post hoc prediction in these other areas, why should it not be favored here? Clearly, greater credence must be given to a theory whose falsifiable, a priori predictions are supported than to a theory that does not predict these findings a priori, and that accounts for them post hoc only because of its unfalsifiable explanatory power.

Furthermore, the assessment of scientific theories depends on many other factors besides the ability to fit data. As philosophers of science often note, theories must also be evaluated on falsifiability, parsimony, the ability to produce provocative hypotheses that push a science forward, the existence of direct evidence for their constructs, freedom from conceptual problems in their apparatus, and integrability with theory in neighboring fields. As we have seen, amodal theories suffer problems in all these regards. They are unfalsifiable, they are not parsimonious, they lack direct support, they suffer conceptual problems such as transduction and symbol grounding, and it is not clear how to integrate them with theory in neighboring fields, such as perception and neuroscience. For all of these reasons, we should view amodal theories with caution and skepticism, and we should be open to alternatives.

1.3. The Current Status of Perceptual Symbol Systems

The reemergence of cognition in the mid-twentieth century did not bring a reemergence of perceptually based cognition. As we have seen, representational schemes moved in a nonperceptual direction. Furthermore, theorists were initially hostile to imbuing modern cognitive theories with any perceptual character whatsoever. When Shepard and Metzler (1971) offered initial evidence for image-like representations in working memory (not long-term memory!), they encountered considerable resistance (e.g., Anderson, 1978; Pylyshyn, 1973, 1981). When Kosslyn (1980) presented his theory of imagery, he argued adamantly that permanent representations in long-term memory are amodal, with perceptual images only existing temporarily in working memory (also see Kosslyn, 1976).

The reasons for this resistance are not entirely clear. One factor could be lingering paranoia arising from the attacks of behaviorists and ordinary language philosophers. Another factor could be more recent criticisms of imagery in philosophy, some of which will be addressed later (e.g., Dennett, 1969; Fodor, 1975; Geach, 1957). Perhaps the most serious factor has been uncharitable characterizations of perceptual cognition that fail to appreciate its potential. Critics often base their attacks on weak formulations of the perceptual approach and underestimate earlier theorists. As a result, perceptual theories of knowledge are widely misunderstood.

Consider some of the more common misunderstandings: Perceptual theories of knowledge are generally believed to contain holistic representations instead of componential representations that exhibit productivity. These theories are widely believed to contain only conscious mental images, not unconscious representations. The representations in these theories are often assumed to arise only in the sensory modalities, not in other modalities of experience, such as proprioception and introspection. These theories are typically viewed as containing only static representations, not dynamic ones. These theories are generally construed as failing to support the propositions that underlie description and interpretation. And these theories are often assumed to include only empirical collections of sense data, not genetically constrained mechanisms.

Careful readings of earlier thinkers, however, indicate that perceptual theories of knowledge often go considerably beyond this simplistic stereotype. Many philosophers, for example, have assumed that perceptual representations are componential and produce representations productively (e.g., Locke, Russell, Price). Many have assumed that unconscious representations, then referred to as "dispositions" and "schemata," produce conscious images (e.g., Locke, Kant, Reid, Price). Many have assumed that images can reflect nonsensory experience, most importantly introspection or "reflection" (e.g., Locke, Hume, Kant, Reid). Many have assumed that images can support the type-token mappings that underlie propositions (e.g., Locke, Reid, Russell, Price). Many have assumed that native mechanisms interpret and organize images (e.g., Kant, Reid). All have assumed that images can be dynamic, not just static, representing events as well as snapshots of time.

As these examples suggest, perceptual theories of knowledge should be judged on the basis of their strongest members, not their weakest. My intent here is to develop a powerful theory of perceptual symbols in the contexts of cognitive science and neuroscience. As we shall see, this type of theory can exhibit the strengths of amodal symbol systems while avoiding their problems.

More and more researchers are developing perceptual theories of cognition. In linguistics, cognitive linguists have made the perceptual character of knowledge a central assumption of their theories (e.g., Fauconnier, 1985, 1997; Jackendoff, 1987; Johnson, 1987; Lakoff, 1987, 1988; Lakoff & Johnson, 1980; Lakoff & Turner, 1989; Langacker, 1986, 1987, 1991, 1997; Sweetser, 1990; Talmy, 1983, 1988; Turner, 1996). In psychology, these researchers include Paivio (1971, 1986), Miller and Johnson-Laird (1976), Huttenlocher (1973, 1976), Shannon (1987), J. Mandler (1992), Tomasello (1992), L. Smith (L. Smith & Heise, 1992; L. Smith, Jones, & Landau, 1992; Jones & L. Smith, 1993), Gibbs (1994), Glenberg (1997), Goldstone (Goldstone, 1994; Goldstone and Barsalou, 1997), Wu (1995), Solomon (1997), MacWhinney (1998), and myself (Barsalou, 1993; Barsalou, Yeh, Luka, Olseth, Mix, & Wu, 1993; Barsalou & Prinz, 1997; Barsalou, Solomon, & Wu, in press). In philosophy, these researchers include Barwise and Etchemendy (1990, 1991), Nersessian (1992), Peacocke (1992), Thagard (1992), Davies and Stone (1995), Heal (1996), Newton (1996), Clark (1997), and Prinz (1997; Prinz and Barsalou, in press). In artificial intelligence, Glasgow (1993) has shown that perceptual representations can increase computational power substantially, and other researchers are grounding machine symbols in sensory-motor events (e.g., Bailey, Feldman, Narayanan, & Lakoff, 1997; Cohen, Atkin, Oates, & Beal, 1997; Rosenstein & Cohen, 1998). Many additional researchers have considered the role of perceptual representations in imagery (e.g., Farah, 1995; Kosslyn, 1994; Finke, 1989; Shepard & Cooper, 1982; Tye, 1991), but the focus here is on perceptual representations in long-term knowledge.

1.4. Recording Systems versus Conceptual Systems

It is widely believed that perceptually based theories of knowledge do not have sufficient expressive power to implement a fully functional conceptual system. As described earlier (1.2.1), a fully functional conceptual system represents both types and tokens, it produces categorical inferences, it combines symbols productively to produce limitless conceptual structures, it produces propositions by binding types to tokens, and it represents abstract concepts. The primary purpose of this target article is to demonstrate that perceptual symbol systems can implement these functions naturally and powerfully.

The distinction between a recording system and a conceptual system is central to this task (Dretske, 1995; Haugeland, 1991). Perceptually based theories of knowledge are typically construed as recording systems. A recording system captures physical information by creating attenuated (not exact) copies of it, as exemplified by photographs, videotapes, and audiotapes. Notably, a recording system does not interpret what each part of a recording contains--it simply creates an attenuated copy. For example, a photo of a picnic simply records the light present at each point in the scene without interpreting the types of entities present.

In contrast, a conceptual system interprets the entities in a recording. In perceiving a picnic, the human conceptual system might construe perceived individuals as instances of tree, table, watermelon, eat, above, and so forth. To accomplish this, the conceptual system binds specific tokens in perception (i.e., individuals) to knowledge for general types of things in memory (i.e., concepts). Clearly, a system that only records perceptual experience cannot construe individuals in this manner--it only records them in the holistic context of an undifferentiated event.

A conceptual system has other properties as well. First, it is inferential, allowing the cognitive system to go beyond perceptual input. Theorists have argued for years that the primary purpose of concepts is to provide categorical inferences about perceived individuals. Again, this is not something that recording systems accomplish. How does a photo of a dog go beyond what it records to provide inferences about the individual present? Second, a conceptual system is productive in the sense of being able to construct complex concepts from simpler ones. This, too, is not something possible with recording systems. How could a photo of some snow combine with a photo of a ball to form the concept of snowball? Third, a conceptual system supports the formulation of propositions, where a proposition results from binding a concept (type) to an individual (token) in a manner that is true or false. Again, this is something that lies beyond the power of recording systems. How does a photo of a dog implement a binding between a concept and an individual?

As long as perceptually based theories of knowledge are viewed as recording systems, they will never be plausible, much less competitive. To become plausible and competitive, a perceptually based theory of knowledge must exhibit the properties of a conceptual system. The primary purpose of this target article is to demonstrate that this is possible.

Of course, it is important to provide empirical support for such a theory as well. Various sources of empirical evidence will be offered throughout the paper, especially in Section 4, and further reports of empirical support are forthcoming (Barsalou et al., in press; Solomon & Barsalou, 1998a,b; Wu & Barsalou, 1998). However, the primary support here will be of a theoretical nature. Because so few theorists currently believe that a perceptually based theory of knowledge could possibly have the requisite theoretical properties, it is essential to demonstrate that it can. Once this has been established, an empirical case can follow.

1.5. Overview

The remainder of this paper presents a theory of perceptual symbols. Section 2 presents six core properties that implement a basic conceptual system: Perceptual symbols are neural representations in sensory-motor areas of the brain (2.1); they represent schematic components of perceptual experience, not entire holistic experiences (2.2); they are multimodal, arising across the sensory modalities, proprioception, and introspection (2.3). Related perceptual symbols become integrated into a simulator that produces limitless simulations of a perceptual component (e.g., red, lift, hungry, 2.4). Frames organize the perceptual symbols within a simulator (2.5), and words associated with simulators provide linguistic control over the construction of simulations (2.6).

Section 3 presents four further properties, derived from the six core properties, that implement a fully functional conceptual system: Simulators can be combined combinatorially and recursively to implement productivity (3.1); they can become bound to perceived individuals to implement propositions (3.2). Because perceptual symbols reside in sensory-motor systems, they exhibit variable embodiment, not functionalism (3.3). Using complex simulations of combined physical and introspective events, perceptual symbol systems represent abstract concepts (3.4).

Section 4 sketches implications of this approach. Viewing knowledge as grounded in sensory-motor areas changes how we think about basic cognitive processes, including categorization, concepts, attention, working memory, long-term memory, language, problem solving, decision making, skill, reasoning, and formal symbol manipulation (4.1). This approach also has implications for evolution and development (4.2), neuroscience (4.3), and artificial intelligence (4.4).

The properties of this theory will not be characterized formally, nor will they be grounded in specific neural mechanisms. Instead, this formulation of the theory should be viewed as a high-level functional account of how the brain could implement a conceptual system using sensory-motor mechanisms. Once the possibility of such an account has been established, later work can develop computational implementations and ground them more precisely in neural systems.

Because this target article focuses on the high level architecture of perceptual symbol systems, it leaves many details unspecified. The theory does not specify the features of perception, nor why attention focuses on some features but not others. The theory does not address how the cognitive system divides the world into categories, nor how abstraction processes establish categorical knowledge. The theory does not explain how the fit between one representation and another is computed, nor how constraints control the combination of concepts. Notably, these issues remain largely unresolved in all theories of knowledge--not just perceptual symbol systems--thereby constituting some of the field's significant challenges. To provide these missing aspects of the theory would exceed the scope of this article, both in length and ambition. Instead, the goal is to formulate the high-level architecture of perceptual symbol systems, which may well provide leverage in resolving these other issues. From hereon, footnotes indicate critical aspects of the theory that remain to be developed.

Finally, this target article proposes a theory of knowledge, not a theory of perception. Although the theory relies heavily on perception, it remains largely agnostic about the nature of perceptual mechanisms. Instead, the critical claim is that whatever mechanisms happen to underlie perception, an important subset will underlie knowledge as well.

2.1. Neural Representations in Sensory-Motor Systems

Perceptual symbols are not like physical pictures; nor are they mental images or any other form of conscious subjective experience. As natural and traditional as it is to think of perceptual symbols in these ways, this is not the form they take here. Instead, they are records of the neural states that underlie perception. During perception, systems of neurons in sensory-motor regions of the brain capture information about perceived events in the environment and in the body. At this level of perceptual analysis, the information represented is relatively qualitative and functional (e.g., the presence or absence of edges, vertices, colors, spatial relations, movements, pain, heat). The neuroscience literature on sensory-motor systems is replete with accounts of this neural architecture (e.g., Bear, Connors, & Paradiso, 1996; Gazzaniga, Ivry, & Mangun, 1998; Zeki, 1993). There is little doubt that the brain uses active configurations of neurons to represent the properties of perceived entities and events.

This basic premise of modern perceptual theory underlies the present theory of perceptual symbol systems: A perceptual symbol is a record of the neural activation that arises during perception. Essentially the same assumption also underlies much current work in imagery: Common neural systems underlie imagery and perception (e.g., Crammond, 1997; Deschaumes-Molinaro, Dittmar, & Vernet-Maury, 1992; Farah, 1995; Jeannerod, 1994, 1995; Kosslyn, 1994; Zatorre, Halpern, Perry, Meyer, & Evans, 1996). The proposal here is stronger, however, further assuming that the neural systems common to imagery and perception underlie conceptual knowledge as well.

This claim by no means implies that identical systems underlie perception, imagery, and knowledge. Obviously, they must differ in important ways. For example, Damasio (1989) suggests that convergence zones integrate information in sensory-motor maps to represent knowledge. More generally, associative areas throughout the brain appear to play this integrative role (Squire, Knowlton, & Musen, 1993). Although mechanisms outside sensory-motor systems enter into conceptual knowledge, perceptual symbols always remain grounded in these systems. Complete transductions never occur whereby amodal representations that lie in associative areas totally replace modal representations. Thus, Damasio (1989) states that convergence zones "are uninformed as to the content of the representations they assist in attempting to reconstruct. The role of convergence zones is to enact formulas for the reconstitution of fragment-based momentary representations of entities or events in sensory and motor cortices" (p. 46).⁵

2.1.1. Conscious versus unconscious processing. Although neural representations define perceptual symbols, they may produce conscious counterparts on some occasions. On other occasions, however, perceptual symbols function unconsciously, as during preconscious processing and automatized skills. Most importantly, the basic definition of perceptual symbols resides at the neural level: Unconscious neural representations--not conscious mental images--constitute the core content of perceptual symbols.⁶

The cognitive and neuroscience literatures support this distinction between unconscious neural representations and optional conscious counterparts. In the cognitive literature, research on preconscious processing indicates that conscious states may not accompany unconscious processing, and that if they do, they follow it (e.g., Marcel, 1983a,b; Velmans, 1991). Similarly, research on skill acquisition has found that conscious awareness falls away as automaticity develops during skill acquisition, leaving unconscious mechanisms largely in control (e.g., Shiffrin, 1988; Shiffrin & Schneider, 1977; Schneider & Shiffrin, 1977). Researchers have similarly found that conscious experience often fails to reflect the unconscious mechanisms controlling behavior (e.g., Nisbett & Wilson, 1977). In the neuroscience literature, research on blindsight indicates that unconscious processing can occur in the absence of conscious visual images (e.g., Cowey & Stoerig, 1991; Weiskrantz, 1986). Similarly, conscious states typically follow unconscious states when processing sensations and initiating actions, rather than preceding them (Dennett & Kinsbourne, 1992; Libet, 1982, 1985). Furthermore, different neural mechanisms appear responsible for producing conscious and unconscious processing (e.g., Farah & Feinberg, 1997; Gazzaniga, 1988; Schacter, McAndrews, & Moscovitch, 1988).

Some individuals experience little or no imagery. By distinguishing unconscious perceptual processing from conscious perceptual experience, we can view such individuals as people whose unconscious perceptual processing underlies cognition with little conscious awareness. If human knowledge is inherently perceptual, there is no a priori reason it must be represented consciously.

2.2. Schematic Perceptual Symbols

A perceptual symbol is not the record of the entire brain state that underlies a perception. Instead, it is only a very small subset that represents a coherent aspect of the state. This is an assumption of many older theories (e.g., Locke, 1690/1959), as well as many current ones (e.g., Langacker, 1986; J. Mandler, 1992; Talmy, 1983). Rather than containing an entire holistic representation of a perceptual brain state, a perceptual symbol contains only a schematic aspect.

The schematic nature of perceptual symbols falls out naturally from two attentional assumptions that are nearly axiomatic in cognitive psychology: Selective attention (a) isolates information in perception, and (b) stores the isolated information in long-term memory. First, consider the role of selective attention in isolating features. During a perceptual experience, the cognitive system can focus attention on a meaningful, coherent aspect of perception. On perceiving an array of objects, attention can focus on the shape of one object, filtering out its color, texture, and position, as well as the surrounding objects. From decades of work on attention, we know that people have a sophisticated and flexible ability to focus attention on features (e.g., Treisman, 1969; Norman, 1976; Shiffrin, 1988), as well as on the relations between features (e.g., Treisman, 1993). Although nonselected information may not be filtered out completely, there is no doubt that it is filtered to a significant extent (e.g., Garner, 1974, 1978; Melara & Marks, 1990).⁷

Once an aspect of perception has been selected, it has a very high likelihood of being stored in long-term memory. On selecting the shape of an object, attention stores information about it. From decades of work on episodic memory, it is clear that where selective attention goes, long-term storage follows, at least to a substantial extent (e.g., Barsalou, 1995; F. Craik & Lockhart, 1972; Morris, Bransford, & Franks, 1977; D. Nelson, Walling, & McEvoy, 1979). Research on the acquisition of automaticity likewise shows that selective attention controls storage (Compton, 1995; Lassaline & Logan, 1993; Logan & Etherton, 1994; Logan, Taylor, & Etherton, 1996). Although some nonselected information may be stored, there is no doubt that it is stored to a much lesser extent than selected information. Because selective attention focuses constantly on aspects of experience in this manner, large numbers of schematic representations become stored in memory. As we shall see later, these representations can serve basic symbolic functions. Section 3.1 demonstrates that these representations combine productively to implement compositionality, and Section 3.2 demonstrates that they acquire semantic interpretations through the construction of propositions. The use of "perceptual symbols" to this point anticipates these later developments of the theory.

Finally, this symbol formation process should be viewed in terms of the neural representations described in Section 2.1. If a configuration of active neurons underlies a perceptual state, selective attention operates on this neural representation, isolating a subset of active neurons. If selective attention focuses on an object's shape, the neurons representing this shape are selected, and a record of their activation is stored. Such storage could reflect the Hebbian strengthening of connections between active neurons (e.g., Pulvermüller, in press), or the indirect integration of active neurons via an adjacent associative area (e.g., Damasio, 1989). Conscious experience may accompany the symbol formation process and may be necessary for this process to occur initially, falling away only as a symbol's processing becomes automatized with practice. Most fundamentally, however, the symbol formation process selects and stores a subset of the active neurons in a perceptual state.

2.2.1. Perceptual symbols are dynamic, not discrete. Once a perceptual symbol is stored, it does not function rigidly as a discrete symbol. Because a perceptual symbol is an associative pattern of neurons, its subsequent activation has dynamical properties. Rather than being reinstated exactly on later occasions, its activations may vary widely. The subsequent storage of additional perceptual symbols in the same association area may alter connections in the original pattern, causing subsequent activations to differ. Different contexts may distort activations of the original pattern, as connections from contextual features bias activation towards some features in the pattern more than others. In these respects, a perceptual symbol is an attractor in a connectionist network. As the network changes over time, the attractor changes. As the context varies, activation of the attractor covaries. Thus, a perceptual symbol is neither rigid nor discrete.

2.2.2. Perceptual symbols are componential, not holistic. Theorists often view perceptual representations as conscious holistic images. This leads to various misunderstandings about perceptual theories of knowledge. One is that it becomes difficult to see how a perceptual representation could be componential. How can one construct a schematic image of a shape without orientation combined holistically? If one imagines a triangle consciously, is orientation not intrinsically required in a holistic image?

It may be true that conscious images must contain certain conjunctions of dimensions. Indeed, it may be difficult or impossible to construct a conscious image that breaks apart certain dimensions, such as shape and orientation. If a perceptual symbol is defined as an unconscious neural representation, however, this is not a problem. The neurons for a particular shape could be active, while no neurons for a particular orientation are. During the unconscious processing of perceptual symbols, the perceptual symbol for a particular shape could represent the shape componentially, while perceptual symbols for other dimensions, such as orientation, remain inactive. The neuroanatomy of vision supports this proposal, given distinct channels in the visual system process different dimensions, such as shape, orientation, color, movement, and so forth.

When conscious images are constructed for a perceptual symbol, the activation of other dimensions may often be required. For example, consciously imagining a triangle may require that it have a particular orientation. However, these conscious representations need not be holistic in the sense of being irreducible to schematic components. For example, Kosslyn and his colleagues have shown that when people construct conscious images, they construct them sequentially, component by component, not holistically in a single step (Kosslyn, Cave, Provost, & von Gierke, 1988; Roth & Kosslyn, 1988; also see Tye, 1993).

2.2.3. Perceptual symbols need not represent specific individuals. Contrary to what some thinkers have argued, perceptual symbols need not represent specific individuals (e.g., Berkeley, 1710/1982; Hume, 1739/1978). Because of the schematicity assumption and its implications for human memory, we should be surprised if the cognitive system ever contains a complete representation of an individual. Furthermore, because of the extensive forgetting and reconstruction that characterize human memory, we should again be surprised if the cognitive system ever remembers an individual with perfect accuracy, during either conscious or unconscious processing. Typically, partial information is retrieved, and some information may be inaccurate.

As we shall see later, the designation of a perceptual symbol determines whether it represents a specific individual or a kind--the resemblance of a symbol to its referent is not critical. Suffice it to say for now that the same perceptual symbol can represent a variety of referents, depending on how causal and contextual factors link it to referents in different contexts (e.g., Dretske, 1995; Fodor, 1975; Goodman, 1976; Schwartz, 1981). Across different pragmatic contexts, a schematic drawing of a generic skyscraper could stand for the Empire State Building, for skyscrapers in general, or for clothing made in New York City. A drawing of the Empire State Building could likewise stand for any of these referents. Just as different physical replicas can stand for each of these referents in different contexts, perceptual representations of them can do so as well (Price, 1953). Thus, the ability of a perceptual symbol to stand for a particular individual need not imply that it must represent an individual.

2.2.4. Perceptual symbols can be indeterminate. Theorists sometimes argue that because perceptual representations are picture-like, they are determinate. It follows that if human conceptualizations are indeterminate, perceptual representations cannot represent them (e.g., Dennett, 1969; but see Block, 1983). For example, it is has been argued that people's conceptualizations of a tiger are indeterminate in its number of stripes; hence they must not be representing it perceptually. To my knowledge, it has not been verified empirically that people's conceptualizations of tigers are in fact indeterminate. If this is true, though, a perceptual representation of a tiger's stripes could be indeterminate in several ways (Schwartz, 1981; Tye, 1993). For example, the stripes could be blurred in an image, such that they are difficult to count. Or, if a perceptual symbol for stripes had been extracted schematically from the perception of a tiger, it might not contain all of the stripes but only a patch. In later representing the tiger, this free-floating symbol might be retrieved to represent the fact that the tiger was striped, but, because it was only a patch, it would not imply a particular number of stripes in the tiger. If this symbol were used to construct stripes on the surface of a simulated tiger, the tiger would then have a determinate number of stripes, but the number might differ from the original tiger, assuming for any number of reasons that the rendering of the tiger's surface did not proceed veridically.

The two solutions considered thus far assume conscious perceptual representations of a tiger. Unconscious neural representations provide another solution. It is well known that high-level neurons in perceptual systems can code information qualitatively. For example, a neuron can code the presence of a line without coding its specific length, position, or orientation. Similarly, a neuron can code the spatial frequency of a region independently of its size or location. Imagine that certain neurons in the visual system respond to stripes independently of their number (i.e., detectors for spatial frequency). In perceiving a tiger, if such detectors fire and become stored in a perceptual representation, they code a tiger's number of stripes indeterminately, because they simply respond to striped patterning and do not capture any particular number of stripes.

Qualitatively oriented neurons provide a perceptual representation system with the potential to represent a wide variety of concepts indeterminately. Consider the representation of triangle. Imagine that certain neurons represent the presence of lines independently of their length, position, and orientation. Further imagine that other neurons represent vertices between pairs of lines independently of the angle between them. Three qualitative detectors for lines, coupled spatially with three qualitative detectors for vertices that join them, could represent a generic triangle. Because all of these detectors are qualitative, the lengths of the lines and the angles between them do not matter; they represent all instances of triangle simultaneously. In this manner, qualitatively specified neurons support perceptual representations that are not only indeterminate but also generic.⁸

2.3. Multimodal Perceptual Symbols

The symbol formation process just described Section 2.2 can operate on any aspect of perceived experience. Not only does it operate on vision, it operates on the other four sensory modalities (audition, haptics, olfaction, and gustation), as well as on proprioception and introspection. In any modality, selective attention focuses on aspects of perceived experience and stores records of them in long-term memory, which later function as symbols. As a result, a wide variety of symbols is stored. From audition, people acquire perceptual symbols for speech and the various sounds heard in everyday experience. From touch, people acquire perceptual symbols for textures and temperatures. From proprioception, people acquire perceptual symbols for hand movements and body positions.

Presumably, each type of symbol becomes established in its respective brain area. Visual symbols become established in visual areas, auditory symbols in auditory areas, proprioceptive symbols in motor areas, and so forth. The neuroscience literature on category localization supports this assumption. When a sensory-motor area is damaged, categories that rely on it during the processing of perceived instances exhibit deficits in conceptual processing (e.g., Damasio & Damasio, 1994; Gainotti et al., 1995; Pulvermüller, in press; Warrington & Shallice, 1984). For example, damage to visual areas disrupts the conceptual processing of categories specified by visual features (e.g., birds). Analogously, damage to motor and somatosensory areas disrupts the conceptual processing of categories specified by motor and somatosensory features (e.g., tools). Recent neuroimaging studies on people with intact brains provide converging evidence (e.g., A. Martin, Haxby, Lalonde, Wiggs, & Ungerleider, 1995; A. Martin, Wiggs, Ungerleider, & Haxby, 1996; Pulvermüller, in press; Rösler, Heil, & Hennighausen, 1995). When normal subjects perform conceptual tasks with animals, visual areas are highly active; when they perform conceptual tasks with tools, motor and somatosensory areas are highly active. Analogous findings have also been found for the conceptual processing of color and space (e.g., DeRenzi & Spinnler, 1967; Levine, Warach, & Farah, 1985; Rösler et al., 1995).

As these findings illustrate, perceptual symbols are multimodal, originating in all modes of perceived experience, and they are distributed widely throughout the modality-specific areas of the brain. It should now be clear that "perceptual" is not being used in its standard sense here. Rather than only referring to the sensory modalities, as it usually does, it refers much more widely to any aspect of perceived experience, including proprioception and introspection.

2.3.1. Introspection. Relative to sensory-motor processing in the brain, introspective processing is poorly understood. Functionally, three types of introspective experience appear especially important: representational states, cognitive operations, and emotional states. Representational states include the representation of an entity or event in its absence, as well as construing a perceived entity as belonging to a category. Cognitive operations include rehearsal, elaboration, search, retrieval, comparison, and transformation. Emotional states include emotions, moods, and affects. In each case, selective attention focuses on an aspect of an introspective state and stores it in memory for later use as a symbol. For example, selective attention could focus on the ability to represent something in its absence, filtering out the particular entity or event represented and storing a schematic representation of a representational state. Similarly, selective attention could focus on the process of comparison, filtering out the particular entities compared and storing a schematic representation of the comparison process. During an emotional event, selective attention could focus on emotional feelings, filtering out the specific circumstances leading to the emotion, and storing a schematic representation of the experience's 'hot' components.

Much remains to be learned about the neural bases of introspection, although much is known about the neural bases of emotion (e.g., Damasio, 1994; LeDoux, 1996). To the extent that introspection requires attention and working memory, the neural systems that underlie them may be central (e.g., Posner, 1995; Rushworth & Owen, 1998; Jonides & E. Smith, 1997). Like sensory-motor systems, introspection may have roots in evolution and genetics. Just as genetically constrained dimensions underlie vision (e.g., color, shape, depth), genetically constrained dimensions may also underlie introspection. Across individuals and cultures, these dimensions may attract selective attention, resulting in the extraction of similar perceptual symbols for introspection across individuals and cultures. Research on mental verbs in psychology and linguistics suggests what some of these dimensions might be (e.g., Cacciari & Levorato, 1994; Levin, 1995; Schwanenflugel, Fabricius, Noyes, Bigler, & Alexander, 1994). For example, Schwanenflugel et al. report that the dimensions of perceptual / conceptual, certain / uncertain, and creative / non-creative organize mental verbs such as see, reason, know, guess, and compare. The fact that the same dimensions arise cross-culturally suggests that different cultures conceptualize introspection similarly (e.g., D'Andrade, 1987; Schwanenflugel, Martin, & Takahashi, 1998).

2.4. Simulators and Simulations

Perceptual symbols do not exist independently of one another in long-term memory. Instead, related symbols become organized into a simulator that allows the cognitive system to construct specific simulations of an entity or event in its absence (analogous to the simulations that underlie mental imagery). Consider the process of storing perceptual symbols while viewing a particular car. As one looks at the car from the side, selective attention focuses on various aspects of its body, such as wheels, doors, and windows. As selective attention focuses on these aspects, the resulting memories are integrated spatially, perhaps using an object-centered reference frame. Similarly, as the perceiver moves to the rear of the car, to the other side, and to the front, stored perceptual records likewise become integrated into this spatially organized system. As the perceiver looks under the hood, peers into the trunk, and climbs inside the passenger area, further records become integrated. As a result of organizing perceptual records spatially, perceivers can later simulate the car in its absence. They can anticipate how the car would look from its side if they were to move around the car in the same direction as before; or they can anticipate how the car would look from the front if they were to around the car in the opposite direction. Because they have integrated the perceptual information extracted earlier into an organized system, they can later simulate coherent experiences of the object.⁹

A similar process allows people to simulate event sequences. Imagine that someone presses the gas pedal and hears the engine roar, then lets up and hears the engine idle. Because the perceptual information stored for each subevent is not stored independently but is instead integrated temporally, the perceiver can later simulate this event sequence. Furthermore, the simulated event may contain multimodal aspects of experience, to the extent that they received selective attention. Besides visual information, the event sequence might include the proprioceptive experience of pressing the pedal, the auditory experience of hearing the engine roar, the haptic experience of feeling the car vibrating, and mild excitement about the power experienced.

As described later (2.5), the perceptual symbols extracted from an entity or event are integrated into a frame that contains perceptual symbols extracted from previous category members. For example, the perceptual symbols extracted from a car are integrated into the frame for car, which contains perceptual symbols extracted from previous instances. After processing many cars, a tremendous amount of multimodal information becomes established that specifies what it is like to experience cars sensorially, proprioceptively, and introspectively. In other words, the frame for car contains extensive multimodal information of what it is like to experience this type of thing.

A frame is never experienced directly in its entirety. Instead, subsets of frame information become active to construct specific simulations in working memory (2.4.3, 2.5.2). For example, a subset of the car frame might become active to simulate one particular experience of a car. On other occasions, different subsets might become active to simulate other experiences. Thus, a simulator contains two levels of structure: (1) An underlying frame that integrates perceptual symbols across category instances, and (2) the potentially infinite set of simulations that can be constructed from the frame. As we shall see in later sections, these two levels of structure support a wide variety of important conceptual functions.¹⁰

2.4.1. Caveats. Several caveats about simulators are essential. First, a simulator produces simulations that are always partial and sketchy, never complete. As selective attention extracts perceptual symbols from perception, it never extracts all of the information potentially available. As a result, a frame is impoverished relative to the perceptions that produced it, as are the simulations constructed from it.

Second, simulations are likely to be biased and distorted in various ways, rarely, if ever, being completely veridical. The well-known principles of Gestalt organization provide good examples. When a linear series of points is presented visually, an underlying line is perceived. As a result, the stored perceptual information goes beyond what is objectively present, representing a line, not just the points. The process of completion similarly goes beyond the information present. When part of an object's edge is occluded by a closer object, the edge is stored as complete, even though the entire edge was not perceived. Finally, when an imperfect edge exists on a perceived object, the perceptual system may idealize and store the edge as perfectly straight, because doing so simplifies processing. As a result, the storage of perceptual symbols may be nonveridical, as may the simulations constructed from them. McCloskey (1983) and Pylyshyn (1978) cite further distortions, and Tye (1993) suggests how perceptual simulation can explain them.

Third, a simulator is not simply an empirical collection of sense impressions but goes considerably further. Mechanisms with strong genetic constraints almost certainly play central roles in establishing, maintaining, and running simulators. For example, genetic predispositions that constrain the processing of space, objects, movement, and emotion underlie the storage of perceptual symbols and guide the simulation process (cf. Baillargeon, 1995; E. Markman, 1989; Spelke, Breinlinger, Macomber, & Jacobson, 1992). Clearly, however, the full-blown realization of these abilities reflects considerable interaction with the environment (e.g., Elman, Bates, Johnson, Karmiloff-Smith, Parisi, & Plunkett, 1996). Thus, a simulator is a both 'rational' and an 'empirical' system, reflecting intertwined genetic and experiential histories.

2.4.2. Dispositions, schemata, and mental models. Simulators bear important similarities with other constructs. In the philosophical literature, Lockean (1690/1959) dispositions and Kantian (1787/1965) schemata are comparable ideas. Both assume that unconscious generative mechanisms produce specific images of entities and events that go beyond particular entities and events experienced in the past. Similar ideas exist in more recent literatures, including Russell (1919b), Price (1953), and Damasio (1994). In all cases, two levels of structure are proposed: A deep set of generating mechanisms produces an infinite set of surface images, with the former typically being unconscious, and the latter typically being conscious.

Mental models are also related to simulators although they are not identical (K. Craik, 1943; Gentner & Stevens, 1983; Johnson-Laird, 1983). Whereas a simulator includes two levels of structure, mental models are roughly equivalent to only the surface level, namely, simulations of specific entities and events. Mental models tend not to address underlying generative mechanisms that produce a family of related simulations.

2.4.3. Concepts, conceptualizations, and categories. According to this theory, the primary goal of human learning is to establish simulators. During childhood, the cognitive system expends much of its resources developing simulators for important types of entities and events. Once individuals can simulate a kind of thing to a culturally acceptable degree, they have an adequate understanding of it. What is deemed a culturally competent grasp of a category may vary, but in general it can be viewed as the ability to simulate the range of multimodal experiences common to the majority of a culture's members (cf. Romney, Weller, & Batchelder, 1986). Thus, people have a culturally acceptable simulator for chair if they can construct multimodal simulations of the chairs typically encountered in their culture, as well as the activities associated with them.

In this theory, a concept is equivalent to a simulator. It is the knowledge and accompanying processes that allow an individual to represent some kind of entity or event adequately. A given simulator can produce limitless simulations of a kind, with each simulation providing a different conceptualization of it. Whereas a concept represents a kind generally, a conceptualization provides one specific way of thinking about it. For example, the simulator for chair can simulate many different chairs under many different circumstantes, each offering a different conceptualization of the category. For further discussion of this distinction between permanent knowledge of a kind in long-term memory and temporary representations of it in working memory, see Barsalou (1987, 1989, 1993; also see 2.4.5).

Simulators do not arise in a vacuum but develop to track meaningful units in the world. As a result, knowledge can accumulate for each unit over time and support optimal interactions with it (e.g., Barsalou et al., 1993; Barsalou, Huttenlocher, & Lamberts, 1998; Millikan, 1998). Meaningful units include important individuals (e.g., family members, friends, personal possessions) and categories (e.g., natural kinds, artifacts, events), where a category is a set of individuals in the environment or introspection. Once a simulator becomes established in memory for a category, it helps identify members of the category on subsequent occasions, and it provides categorical inferences about them, as described next.¹¹

2.4.4. Categorization, categorical inferences, and affordances. Tracking a category successfully requires that its members be categorized correctly when they appear. Viewing concepts as simulators suggests a different way of thinking about categorization. Whereas many theories assume that relatively static, amodal structures determine category membership (e.g., definitions, prototypes, exemplars, theories), simulators suggest a more dynamic, embodied approach: If the simulator for a category can produce a satisfactory simulation of a perceived entity, the entity belongs in the category. If the simulator cannot produce a satisfactory simulation, the entity is not a category member.¹²

Besides being dynamic, grounding categorization in perceptual symbols has another important feature: The knowledge that determines categorization is represented in roughly the same manner as the perceived entities that must be categorized. For example, the perceptual simulations used to categorize chairs approximate the actual perceptions of chairs. In contrast, amodal theories assume that amodal features in concepts are compared to perceived entities to perform categorization. Whereas amodal theories have to explain how two very different types of representation are compared, perceptual symbol systems simply assume that two similar representations are compared. As a natural side effect of perception, perceptual knowledge accures that can be compared directly to perceived entities during categorization.

On this view, categorization depends on both familiar and novel simulations. Each successful categorization stores a simulation of the entity categorized. If the same entity or a highly similar entity is encountered later, it is assigned to the category because the perception of it matches an existing simulation in memory. Alternatively, if a novel entity is encountered that fails to match an existing simulation, constructing a novel simulation that matches the entity can establish membership. Explanation-based learning assumes a similar distinction between expertise and creativity in categorization (DeJong & Mooney, 1986; T. Mitchell, Keller, & Kedar-Cabelli, 1986), as do theories of skill acquisition (Anderson, 1993; Logan, 1988; Newell, 1990), although these approaches typically adopt amodal representations.

As an example, imagine that the simulator for triangle constructs three lines and connects their ends uniquely. Following experiences with previous triangles, simulations that match these instances become stored in the simulator. On encountering these triangles later, or highly similar ones, prestored simulations support rapid categorization, thereby implementing expertise. However, a very different triangle never seen before can also be categorized if the triangle simulator can construct a simulation of it (cf. Miller & Johnson-Laird, 1976).

Categorization is not an end in itself but provides access to categorical inferences. Once an entity is categorized, knowledge associated with the category provides predictions about the entity's structure, history, and behavior, and also suggests ways of interacting with it (e.g., Barsalou, 1991; Ross, 1996; also see 3.2.2). In this theory, categorical inferences arise through simulation. Because a simulator contains a tremendous amount of multimodal knowledge about a category, it can simulate information that goes beyond the information perceived in a categorized entity. On perceiving a computer from the front, the simulator for computer can simulate all sorts of things not perceived, such as the computer's rear panel and internal components, what the computer will do when turned on, what tasks it can perform, how the keys will feel when pressed, and so forth. Rather than having to learn about the entity from scratch, a perceiver can run simulations that anticipate the entity's structure and behavior, and that suggest ways of interacting with it successfully.

Simulators also produce categorical inferences in the absence of category members. As described later, simulations provide people with a powerful ability to reason about entities and events in their absence (2.6, 3.1, 3.2, 4.1, 4.2). Simulations of future entities, such as a rental home, allow people to identify preparations that must be made in advance. Simulations of future events, such as asking a favor, allow people to identify optimal strategies for achieving success. To the extent that future category members are similar to previous category members, simulations of previous members provide reasonable inferences about future members.¹³

Deriving categorical inferences successfully requires that simulations preserve at least some of the affordances present in actual sensory-motor experiences with category members (cf. Gibson, 1979; also see S. Edelman, in press). To the extent that simulations capture affordances from perception and action, successful reasoning about physical situations can proceed in their absence (Glenberg, 1997; Glenberg et al., 1998b; Newton, 1996). Agents can draw inferences that go beyond perceived entities, and they can plan intelligently for the future. While sitting in a restaurant and wanting to hide from someone entering, one could simulate that a newspaper on the table affords covering one's face completely but that a matchbook does not. As a result of these simulations, the newspaper is selected to achieve this goal rather than the matchbook. Because the simulations captured the physical affordances correctly, the selected strategy works.

2.4.5. Concept stability. Equating concepts with simulators provides a solution to the problem of concept stability. Previous work demonstrates that conceptualizations of a category vary widely between and within individuals (Barsalou, 1987, 1989, 1993). If different people conceptualize bird differently on a given occasion, and if the same individual conceptualizes bird differently across occasions, how can stability be achieved for this concept?

One solution is to assume that a common simulator for bird underlies these different conceptualizations, both between and within individuals. First consider how a simulator produces stability within an individual. If a person's different simulations of a category arise from the same simulator, then they can all be viewed as instantiating the same concept. Because the same simulator produced all of these simulations, it unifies them. Between individuals, the key issue concerns whether different people acquire similar simulators. A number of factors suggest that they should, including a common cognitive system, common experience with the physical world, and socio-cultural institutions that induce conventions (e.g., Newton, 1996; Tomasello, Kruger, & Ratner, 1993). Although two individuals may represent the same category differently on a given occasion, each may have the ability to simulate the other's conceptualization. In an unpublished study, subjects almost always viewed other subjects' conceptualizations of a category as correct, even though their individual conceptualizations varied widely. Each subject produced a unique conceptualization but accepted those of other subjects because they could be simulated. Furthermore, common contextual constraints during communication often drive two people's simulations of a category into similar forms. In another unpublished study, conceptualizations of a category became much more stable both between and within subjects when constructed in a common context. Subjects shared similar simulators that produced similar conceptualizations when constrained adequately.

2.4.6. Cognitive penetration. The notion of a simulator is difficult to reconcile with the view that cognition does not penetrate perception (Fodor, 1983). According to the impenetrability hypothesis, the amodal symbol system underlying higher cognition has little or no impact on processing in sensory-motor systems, because these systems are modular and therefore impenetrable. In contrast, the construct of a simulator assumes that sensory-motor systems are deeply penetrable. Because perceptual symbols reside in sensory-motor systems, running a simulator involves a partial running of these systems in a top-down manner.

In an insightful BBS review of top-down effects in vision, Pylyshyn (in press) concludes that cognition only produces top-down effects indirectly through attention and decision making--it does not affect the content of vision directly. Contrary to this conclusion, however, much evidence indicates that cognition does affect the content of sensory-motor systems directly. The neuroscience literature on mental imagery demonstrates clearly that cognition establishes content in sensory-motor systems in the absence of physical input. In visual imagery, the primary visual cortex, V1, is often active, along with many other early visual areas (e.g., Kosslyn,, Thompson, Kim, & Alpert, 1995). In motor imagery, the primary motor cortex, M1, is often active, along with many other early motor areas (e.g., Crammond, 1997; Deschaumes-Molinaro et al, 1992; Jeannerod, 1994, 1995). Indeed, motor imagery not only activates early motor areas, it also stimulates spinal neurons, produces limb movements, and modulates both respiration and heart rate. When sharpshooters imagine shooting a gun, their entire body behaves similarly to actually doing so. In auditory imagery, activation has not yet been observed in the primary auditory cortex, but activation has been observed in other early auditory areas (e.g., Zatorre et al., 1996). These findings clearly demonstrate that cognition establishes content in sensory-motor systems in the absence of physical input.

A potential response is that mental imagery arises solely within sensory-motor areas--it is not initiated by cognitive areas. In this vein, Pylyshyn (in press) suggests that perceptual modules contain local memory areas that affect the content of perception in a top-down manner. This move undermines the impenetrability thesis, however, at least in its strong form. As a quick perusal of textbooks on cognition and perception reveals, memory is widely viewed as a basic cognitive process--not as a perceptual process. Many researchers would probably agree that once memory is imported into a sensory-motor system, cognition has been imported. Furthermore, to distinguish perceptual memory from cognitive memory, as Pylyshyn does, makes sense only if one assumes that cognition utilizes amodal symbols. Once one adopts the perspective of perceptual symbol systems, there is only perceptual memory, and it constitutes the representational core of cognition. In this spirit, Damasio (1989) argues eloquently that there is no sharp discontinuity between perceptual and cognitive memory. Instead, there is simply a gradient from posterior to anterior association areas in the complexity and specificity of the memories that they activate in sensory-motor areas. On Damasio's view, memory areas both inside and outside a sensory-motor system control its feature map to implement cognitive representations. In this spirit, the remainder of this target article assumes that top-down processing includes all memory effects on perceptual content, including memory effects that originate in local association areas.

Nevertheless, Pylyshyn (in press) makes compelling arguments about the resiliency of bottom-up information in face-to-face competition with contradicting top-down information. For example, when staring at the Müller-Lyer illusion, one cannot perceive the horizontal lines as equivalent in length, even though one knows cognitively that they are. Rather than indicating impenetrability, however, this important observation may simply indicate that bottom-up information dominates top-down information when they conflict (except in the degenerate case of psychosis and other hallucinatory states, when top-down information does dominate bottom-up information). Indeed, Marslen-Wilson and Tyler (1980), while taking a nonmodular interactive approach, offer exactly this account of bottom-up dominance in speech recognition. Although top-down processing can penetrate speech processing, it is overridden when bottom-up information conflicts. If semantic and syntactic knowledge predict that "The cowboy climbed into the _____" ends with "saddle," but the final word is actually "jacuzzi", then "jacuzzi" overrides "saddle".

On this view, sensory-motor systems are penetrable but not always. When bottom-up information conflicts with top-down information, the former usually dominates. When bottom-up information is absent, however, top-down information penetrates, as in mental imagery. Perhaps most critically, when bottom-up and top-down information are compatible, top-down processing again penetrates, but in subtle manners that complement bottom-up processing. The next section (2.4.7) reviews several important phenomena in which bottom-up and top-down processing simultaneously activate sensory-motor representations as they cooperate to perceive physical entities (i.e., implicit memory, filling-in, anticipation, interpretation). Recent work on simultaneous imagery and perception shows clearly that these two processes work well together when compatible (e.g., Craver-Lemley & Reeves, 1997; Gilden, Blake, & Hurst, 1995).

Perhaps the critical issue in this debate concerns the definition of cognition. On Pylyshyn's view, cognition concerns semantic beliefs about the external world (i.e., the belief that the horizontal lines are the same length in the Müller-Lyer illusion). However, this is a far narrower view of cognition than most cognitive psychologists take, as evidenced by journals and texts in the field. Judging from these sources, cognitive psychologists believe that a much broader array of processes and representations--including memory--constitutes cognition.

Ultimately, as Pylyshyn suggests, identifying the mechanisms that underlie intelligence should be our primary goal, from the most preliminary sensory processes to the most abstract thought processes. Where we actually draw the line between perception and cognition may not be all that important, useful, or meaningful. In this spirit, perceptual symbol systems attempt to characterize the mechanisms that underlie the human conceptual system. As we have seen, the primary thesis is that sensory-motor systems represent not only perceived entities but also conceptualizations of them in their absence. From this perspective, cognitive penetrates perception when sensory input is absent, or when top-down inferences are compatible with sensory input.

2.4.7. A family of representational processes. Evolution often capitalizes on existing mechanisms to perform new functions (Gould, 1991). Representational mechanisms in sensory-motor regions of the brain may be such mechanisms. Thus far, these representational mechanisms have played three roles in perceptual symbol systems: (1) In perception, they represent physical objects. (2) In imagery, they represent objects in their absence. (3) In conception, they also represent objects in their absence. On this view, conception differs from imagery primarily in the consciousness and specificity of sensory-motor representations, with these representations being more conscious and detailed in imagery than in conception (Solomon & Barsalou, 1998b; Wu & Barsalou, 1998). Several other cognitive processes also appear to use these same representational mechanisms, including implicit memory, filling-in, anticipation, and interpretation. Whereas perception, imagery, and conception perform either bottom-up or top-down processing exclusively, these other four processes fuse complementary mixtures of bottom-up and top-down processing to construct perceptions.

In implicit memory (i.e., repetition priming), a perceptual memory speeds the perception of a familiar entity (e.g., Roediger & McDermott, 1993; Schacter, 1995). On seeing a particular chair, for example, a memory is established that speeds perception of the same chair later. Much research demonstrates the strong perceptual character of these memories, with slight deviations in perceptual features eliminating facilitory effects. Furthermore, imagining an entity produces much the same facilitation as perceiving it, suggesting a common representational basis. Perhaps most critically, implicit memory has been localized in sensory-motor areas of the brain, with decreasing brain activity required to perceive a familiar entity (e.g., Buckner, Petersen, Ojemann, Miezin, Squire, & Raichle, 1995). Thus, the representations that underlie implicit memory reside in the same systems that process entities perceptually. When a familiar entity is perceived, implicit memories become fused with bottom-up information to represent it efficiently.

In filling-in, a perceptual memory completes gaps in bottom-up information. Some filling-in phenomena reflect perceptual inferences that are largely independent of memory (for a review, see Pessoa, Thompson, & Noë, in press). In the perception of illusory contours, for example, low-level sensory mechanisms infer edges on the basis of perceived vertices (e.g., Kanizsa, 1979). However, other filling-in phenomena rely heavily on memory. In the phoneme restoration effect, knowledge of a word creates the conscious perceptual experience of hearing a phoneme where noise exists physically (e.g., Samuel, 1981, 1987; Warren, 1970). More significantly, phoneme restoration adapts low-level feature detectors much as if physical phonemes had adapted them (Samuel, 1997). Thus, when a word is recognized, its memory representation fills in missing phonemes, not only in experience, but also in sensory processing. Such findings strongly suggest that cognitive and perceptual representations reside in a common system, and that they become fused to produce perceptual representations. Knowledge-based filling-in also occurs in vision. For example, knowledge about bodily movements causes apparent motion to deviate from the perceptual principle of minimal distance (Shiffrar & Freyd, 1990, 1993). Rather than filling in an arm as taking the shortest path through a torso, perceivers fill it in as taking the longer path around the torso, consistent with bodily experience.

In perceptual anticipation, the cognitive system uses past experience to simulate a perceived entity's future activity. For example, if an object traveling along a trajectory disappears, perceivers anticipate where it would be if it were still on the trajectory, recognizing it faster at this point than at the point it disappeared, or at any other point in the display (Freyd, 1987). Recent findings indicate that knowledge affects the simulation of these trajectories. When subjects believe that an ambiguous object is a rocket, they simulate a different trajectory compared to when they believe it is a steeple (Reed & Vinson, 1996). Even infants produce perceptual anticipations in various occlusion tasks (e.g., Baillargeon, 1995; Hespos & Rochat, 1997). These results further indicate that top-down and bottom-up processes coordinate the construction of useful perceptions.

In interpretation, the conceptual representation of an ambiguous perceptual stimulus biases sensory processing. In audition, when subjects believe that multiple speakers are producing a series of speech sounds, they normalize the sounds associated with each speaker (Magnuson & Nusbaum, 1993). In contrast, when subjects believe that only one speaker is producing these same sounds, they do not normalize them, treating them instead as differences in the speaker's emphasis. Thus, each interpretation produces sensory processing that is appropriate for its particular conceptualization of the world. Again, cognition and sensation coordinate to produce meaningful perceptions (also see Nusbaum & Morin, 1992; Schwab, 1981). Analogous interpretive effects occur in vision. Conceptual interpretations guide computations of figure and ground in early visual processing (Peterson & Gibson, 1993; Weisstein & Wong, 1986); they affect the selective adaptation of spatial frequency detectors (Weisstein, Montalvo, & Ozog, 1972; Weisstein & Harris, 1980); and they facilitate edge detection (Weisstein & Harris, 1974). Frith and Dolan (1997) report that top-down interpretive processing activates sensory-motor regions in the brain.

In summary, an important family of basic cognitive processes appears to utilize a single mechanism, namely, sensory-motor representations. These processes, although related, vary along a continuum of bottom-up to top-down processing. At one extreme, bottom-up input activates sensory-motor representations in the absence of top-down processing ('pure' perception). At the other extreme, top-down processing activates sensory-motor representations in the absence of bottom-up processing (imagery and conception). In between lie processes that fuse complementary mixtures of bottom-up and top-down processing to coordinate the perception of physical entities (implicit memory, filling-in, anticipation, interpretation).

2.5. Frames

A frame is an integrated system of perceptual symbols that is used to construct specific simulations of a category. Together, a frame and the simulations it produces constitute a simulator. In most theories, frames are amodal, as are the closely related constructs of schemata and scripts (e.g., Minsky, 1977, 1985; Rumelhart & Ortony, 1978; Schank & Abelson, 1977; for reviews, see Barsalou, 1992; Barsalou & Hale, 1992). As we shall see, however, frames and schemata have natural analogues in perceptual symbol systems. In the account that follows, all aspects require much further development, and many important issues are not addressed. This account is only meant to provide a rough sense of how frames develop in perceptual symbol systems, and how they produce specific simulations.

The partial frame for car in Figure 3 illustrates how perceptual symbol systems implement frames. On the perception of a first car (Figure 3a), the schematic symbol formation process in 2.2 extracts perceptual symbols for the car's overall shape and some of its components, and then integrates these symbols into an object-centered reference frame. The representation at the top of Figure 3a represents the approximate volumetric extent of the car, as well as embedded subregions that contain significant components (e.g., the doors and tires). These subregions and their specializations reflect the result of the symbol formation process described earlier. For every subregion that receives selective attention, an approximate delineation of the subregion is stored and then connected to the content information that specializes it. For example, attending to the window and handle of the front door establishes perceptual symbols for these subregions and the content information that specializes them.¹⁴

As Figure 3a illustrates, the frame represents spatial and content information separately. At one level, the volumetric regions of the object are represented according to their spatial layout. At another level, the contents of these subregions are represented as specializations. This distinction between levels of representation follows the work of Ungerlieder and Mishkin (1982), who identified separate neural pathways for spatial / motor information and object features (also see Milner & Goodale, 1995). Whereas the spatial representation establishes the frame's skeleton, the content specializations flesh it out.

On the perception of a second car, a reminding takes place, retrieving the spatially integrated set of symbols for the first car (Barsalou et al., 1998; Medin & Ross, 1989; Millikan, 1998; Ross, Perkins, & Tenpenny, 1990; Spalding & Ross, 1994). The retrieved set of symbols guides processing of the second car in a top-down manner, leading to the extraction of perceptual symbols in the same subregions. As Figure 3b illustrates, this might lead to the extraction of content information for the second car's shape, doors, and wheels, which become connected to the same subregions as the content extracted from the first car. In addition, other subregions of the second car may be processed, establishing new perceptual symbols (e.g., the antenna and gas cap).

Most important, all the information extracted from the two cars becomes integrated into a knowledge structure that constitutes the beginnings of a car frame. When perceptual symbols have been extracted from the same volumetric region for both cars, they both become connected to that subregion in the object-oriented reference frame. For example, the perceptual symbols for each car's doors become associated with the door subregions. As a result, the specialization from either car can be retrieved to specialize a region during a simulation, as may an average superimposition of them.

Figure 3b further illustrates how connections within the frame change as new instances become encoded. First, specializations from the same instance are connected together, providing a mechanism for later reinstating it. Second, these connections become weaker over time, with the dashed connections for the first instance representing weaker connections than those for the second instance. Third, inhibitory connections develop between specializations that compete for the same subregion, providing an additional mechanism for reinstating particular instances. Finally, connections processed repeatedly become stronger, such as the thicker connections between the overall volume and the subregions for the doors and wheels.

Following experiences with many cars, the car frame accrues a tremendous amount of information. For any given subregion, many specializations exist, and default specializations develop, either through the averaging of specializations, or because one specialization develops the strongest association to the subregion. Frames develop similarly for event concepts (e.g., eating), except that subregions exist over time as well as in space (examples will be presented in Section 3.4 on abstract concepts).

2.5.1. Basic frame properties. Barsalou (1992) and Barsalou and Hale (1992) propose that four basic properties constitute a frame (and the related construct of a schema): (a) predicates, (b) attribute-value bindings, (c) constraints, and (d) recursion. Barsalou (1993) describes how these properties manifest themselves in perceptual symbol systems.

Predicates are roughly equivalent to unspecialized frames. For example, the predicate CAR(door=x, window=y, ...) is roughly equivalent to the perceptual frame for car with its subregions unspecialized (e.g., the hierarchical volumetric representation in Figure 3b).

Attribute-value bindings arise through the specialization of a subregion in a simulation. As different specializations become bound to the same subregion, they establish different 'values' of an 'attribute' or 'slot' (e.g., door and its specializations in Figure 3b).

Constraints arise from associative connections between specializations that represent individuals and subcategories within a frame (e.g., the connections between the specializations for the second instance of car in Figure 3b). Thus, activating the second car's specialization in one subregion activates its associated specializations in other subregions, thereby simulating this specific car. Because constraints can weaken over time, and because strong defaults may dominate the activation process, reconstructive error can occur. For example, if the first car's tires received extensive processing, the strength of their connections to the subregions for the wheels might dominate, even during attempts to simulate the second car.

Recursion arises from establishing simulators within an existing simulator. As Figures 3a and 3b suggest, the simulator for wheel might initially just simulate schematic circular volumes, with all other information filtered out. As more attention is paid to the wheels on later occasions, however, detail is extracted from their subregions. As a result, simulators for tire and hubcap develop within the simulator for wheel, organizing the various specializations that occur for each of them. Because this process can continue indefinitely, an arbitrarily deep system of simulators develops, bounded only by the perceiver's motivation and perceptual resolution.

2.5.2. Constructing specific simulations. The cognitive system uses the frame for a category to construct specific simulations (i.e., mental models). As discussed in later sections, such simulations take myriad forms. In general, however, the process takes the form illustrated in Figure 3c. First, the overall volumetric representation for car becomes active, along with a subset of its subregions (e.g., for doors, wheels, antenna, gas cap). If a cursory simulation is being constructed, only the subregions most frequently processed previously are included, with other subregions remaining inactive. This process can then proceed recursively, with further subregions and specializations competing for inclusion. As Figure 3c illustrates, subregions for the doors and windows subsequently become active, followed by their specializations.

In any given subregion, the specialization having the highest association with the subregion, with other active regions, and with other active specializations becomes active. Because this process occurs for each subregion simultaneously, it is highly interactive. The simulation that emerges reflects the strongest attractor in the frame's state space. If context 'clamps' certain aspects of the simulation initially, the constraint satisfaction process may diverge from the frame's strongest attractor toward a weaker one. If an event is being simulated, subregions and their specializations may change over time as the constraint satisfaction process evolves recurrently.

During a simulation, processing is not limited to the retrieval of frame information but can also include transformations of it. Retrieved information can be enlarged, shrunk, stretched, and reshaped; it can be translated across the simulation spatially or temporally; it can be rotated in any dimension; it can remain fixed while the perspective on it varies; it can be broken into pieces; it can be merged with other information. Other transformations are no doubt possible as well. The imagery literature offers compelling evidence that such transformations are readily available in the cognitive system (e.g., Finke, 1989; Kosslyn, 1980; Shepard & Cooper, 1982), and that these transformations conform closely to perceptual experience (e.g., Freyd, 1987; Parsons, 1987a,b; Shiffrar & Freyd, 1990, 1993).

2.5.3. Framing and background-dependent meaning. As linguists and philosophers have noted for some time, concepts are often specified relative to one another (e.g., Fillmore, 1985; Langacker, 1986; A. Lehrer, 1974; A. Lehrer & Kittay, 1992; Quine, 1953). Psychologists often make a similar observation that concepts are specified in the context of intuitive theories (e.g., Carey, 1985; Keil, 1989; Murphy & Medin, 1985; Rips, 1989; Wellman & Gelman, 1988). Framing and background dependent meaning are two specific ways in which background knowledge specifies concepts. In framing, a focal concept depends on a background concept and cannot be specified independently of it. For example, payment is specified relative to buy, hypotenuse is specified relative to right triangle, and foot is specified relative to leg and body. In background-dependent meaning, the same focal concept changes as its background concept changes. For example, the conceptualization of foot varies as the background concept changes from human to horse to tree. Similarly, the conceptualization of handle varies across shovel, drawer, and car door, as does the conceptualization of red across fire truck, brick, hair, and wine (Halff, Ortony, & Anderson, 1976).

Frames provide the background structures that support framing. Thus, the event frame for buy organizes the background knowledge necessary for understanding payment, and the entity frame for human organizes the background knowledge necessary for understanding foot. When payment or foot is conceptualized, its associated frame produces a simulation that provides the necessary background for understanding it. For example, a simulation of a human body provides one possible framing for a simulation of a foot.

Frames also offer a natural account of background dependent meaning. Foot, for example, is conceptualized differently when human is simulated in the background than when horse or tree is simulated. Because different perceptual symbols are accessed for foot in the context of different frames, simulations of foot vary widely. Similarly, different conceptualizations of red reflect different perceptual symbols accessed in frames for fire truck, brick, hair, and wine.¹⁵

2.6. Linguistic Indexing and Control

In humans, linguistic symbols develop together with their associated perceptual symbols. Like a perceptual symbol, a linguistic symbol is a schematic memory of a perceived event, where the perceived event is a spoken or a written word. A linguistic symbol is not an amodal symbol, nor does an amodal symbol ever develop in conjunction with it. Instead, a linguistic symbol develops just like a perceptual symbol. As selective attention focuses on spoken and written words, schematic memories extracted from perceptual states become integrated into simulators that later produce simulations of these words in recognition, imagination and production.

As simulators for words develop in memory, they become associated with simulators for the entities and events to which they refer. Whereas some simulators for words become linked to simulators for entire entities or events, others become linked to subregions and specializations. Whereas "car" becomes linked to the entire simulator for car, "trunk" becomes linked to one of its subregions. Simulators for words also become associated with other aspects of simulations, including surface properties (e.g., "red"), manners (e.g., "rapidly"), relations (e.g., "above"), and so forth. Within the simulator for a concept, large numbers of simulators for words become associated with its various aspects to produce a semantic field that mirrors the underlying conceptual field (Barsalou, 1991, 1992, 1993).

Once simulators for words become linked to simulators for concepts, they can control simulations. On recognizing a word, the cognitive system activates the simulator for the associated concept to simulate a possible referent. On parsing the sentences in a text, surface syntax provides instructions for building perceptual simulations (Langacker, 1986, 1987, 1991, 1997). As discussed in the next section, the productive nature of language, coupled with the links between linguistic and perceptual simulators, provides a powerful means of constructing simulations that go far beyond an individual's experience. As people hear or read a text, they use productively formulated sentences to construct a productively formulated simulation that constitutes a semantic interpretation (4.1.6). Conversely, during language production, the construction of a simulation activates associated words and syntactic patterns, which become candidates for spoken sentences designed to produce a similar simulation in a listener. Thus, linguistic symbols index and control simulations to provide humans with a conceptual ability that is probably the most powerful of any species (Donald, 1991, 1993). As MacWhinney (1998) suggests, language allows conversationalists to coordinate simulations from a wide variety of useful perspectives.¹⁶

By parsing perception into schematic components, and then integrating components across individuals into frames, simulators develop that represent the types of entities and events in experience. The result is a basic conceptual system, not a recording system. Rather than recording holistic representations of perceptual experience, this system establishes knowledge about the categories of individuals that constitute it. Each simulator represents a type, not a token, and the collection of simulators constitutes a basic conceptual system that represents the components of perception, and provides categorical inferences about them.

We have yet to characterize a fully functional conceptual system. To do so, we must establish that perceptual symbol systems can implement productivity, propositions, and abstract concepts. The next section demonstrates that these properties follow naturally from the basic system presented thus far, and that one additional property--variable embodiment--follows as well.

3.1. Productivity

One of the important lessons we have learned from amodal symbol systems is that a viable theory of knowledge must be productive (e.g., Chomsky, 1957; Fodor, 1975; Fodor & Pylyshyn, 1988). The human cognitive system can produce an infinite number of conceptual and linguistic structures that go far beyond those experienced. No one has experienced a real Cheshire Cat, but it is easy to imagine a cat whose body fades and reappears while its human smile remains.

Productivity is the ability to construct an unlimited number of complex representations from a finite number of symbols using combinatorial and recursive mechanisms. It is fair to say that productivity is not typically recognized as possible within perceptual theories of knowledge, again because these theories are usually construed as recording systems. It is worth noting, however, that certain physical artifacts exhibit a kind of 'perceptual productivity,' using combinatorial and recursive procedures on schematic diagrams to construct limitless complex diagrams. In architecture, notations exist for combining primitive schematic diagrams combinatorially and recursively to form complex diagrams of buildings (e.g., W. Mitchell, 1990). In electronics, similar notations exist for building complex circuits from primitive schematic components (Haugeland, 1991). The fact that physical artifacts can function in this manner suggests that schematic perceptual representations in cognition could behave similarly (Price, 1953). As we shall see, if one adopts the core properties of perceptual symbol systems established thus far, productivity follows naturally.

Figure 4 illustrates productivity in perceptual symbol systems. Before exploring productivity in detail, it is first necessary to make several points about the diagrams in Figure 4, as well as those in all later figures. First, diagrams such as the balloon in Figure 4a should not be viewed as literally representing pictures or conscious images. Instead, these theoretical illustrations stand for configurations of neurons that become active in representing the physical information conveyed in these drawings. For example, the balloon in Figure 4a is a theoretical notation that refers to configurations of neurons active in perceiving balloons.

Second, each of these drawings stands metonymically for a simulator. Rather than standing for only one projection of an object (e.g., the balloon in Figure 4a), these drawings stand for a simulator capable of producing countless projections of the instance shown, as well as of an unlimited number of other instances.

Third, the diagrams in Figure 4b stand for simulators of spatial relations that result from the symbol formation process described earlier. During the perception of a balloon above a cloud, for example, selective attention focuses on the occupied regions of space, filtering out the entities in them. As a result, a schematic representation of above develops that contains two schematic regions of space within one of several possible reference frames (not shown in Figure 4). Following the similar extraction of information on other occasions, a simulator develops that can render many different above relations. For example, specific simulations may vary in the vertical distance between the two regions, in their horizontal offset, and so forth. Finally, the thicker boundary for a given region in a spatial relation indicates that selective attention is focusing on it.¹⁷ Thus, above and below involve the same representation of space, but with a different distribution of attention over it. As much research illustrates, an adequate treatment of spatial concepts requires further analysis and detail than provided here (e.g., Herskovits, 1997; Regier, 1996; Talmy, 1983). Nevertheless, one can view spatial concepts as simulators that develop through the schematic symbol formation process in Section 2.2.

Figure 4c illustrates how the simulators in Figures 4a and 4b combine to produce complex perceptual simulations combinatorially. In the left-most example, the simulator for above produces a specific simulation of above (i.e., two schematic regions of the same size, close together, and in vertical alignment). The simulators for balloon and cloud produce specific simulations that specialize the two regions of the above simulation. The result is a complex simulation in which a balloon is simulated above a cloud. Note that complex simulations of these three categories could take infinitely many forms, including different balloons, clouds, and above relations. The second and third examples in Figure 4c illustrate the combinatorial nature of these simulations, as the various objects in Figure 4a are rotated through each of the regions in above. Because many possible objects could enter into simulations of above, a very large number of such simulations is possible.

A simulation can be constructed recursively by specializing the specialization of a schematic region. In Figure 4d, the lower region of above is specialized recursively with a simulation of left-of, whose regions are then specialized with simulations of jet and cloud. The resulting simulation represents a balloon that is above a jet to the left of a cloud. Because such recursion can occur indefinitely, an infinite number of simulations can be produced in principle.

3.1.1. Reversing the symbol formation process. Productivity in perceptual symbol systems is approximately the symbol formation process run in reverse. During symbol formation, large amounts of information are filtered out of perceptual representations to form a schematic representation of a selected aspect (2.2). During productivity, small amounts of the information filtered out are added back. Thus, schematicity makes productivity possible. For example, if a perceptual symbol for ball only represents its shape schematically, after color and texture have been filtered out, then information about color and texture can later be added productively. For example, the simulation of a ball could evolve into a blue ball or a smooth yellow ball. Because the symbol formation process similarly establishes schematic representations for colors and textures, these representations can be combined productively with perceptual representations for shapes to produce complex simulations.

Productivity is not limited to filling in schematic regions but can also result from replacements, transformations, and deletions of existing structure. Imagine that a simulated lamp includes a white cylindrical shade. To represent different lamps, the simulated shade could be replaced with a simulated cardboard box, it could be transformed to a cone shape, or it could be deleted altogether. Such operations appear widely available for constructing simulations, extending productivity further.

Most important, the complementarity of schematization and specialization allow perceptual systems to go beyond a recording system and become a conceptual system. Whereas photos and videos only capture information holistically, a perceptual symbol system extracts particular parts of images schematically and integrates them into simulators. Once simulators exist, they can be combined to construct simulations productively. Such abilities go far beyond the recording abilities of photos and videos; yet, as we have seen, they can be achieved within a perceptual framework.

3.1.2. Productivity in imagination. Productivity can surpass experience in many ways, constituting an important source of creativity (Barsalou & Prinz, 1997). For example, one can simulate a chair never encountered, such as a pitted lavender chair. Because perceptual symbols for colors become organized together in a semantic field, as do perceptual symbols for textures, simulations of a chair can cycle through the symbols within each field to try out various combinations (Barsalou, 1991, 1992, 1993). During interior decoration, one can combine colors and textures productively with a schematic representation of a chair to see which works best, with the semantic fields for colors and textures providing 'palettes' for constructing the possibilities. Because different semantic fields can be explored orthogonally, the resulting simulations take on an analytic combinatorial character.

Productive mechanisms can further construct simulations that violate properties of the physical world. For example, one can productively combine the shape of a chair with the simulation of a dog to construct a dog that functions as a chair. By searching through the combinatorial space of possibilities, one can construct many similar simulations, such as Carroll' s (1960) flamingos in Alice in Wonderland, who form themselves into croquet mallets, and his hedgehogs who form themselves into croquet balls. The space of such possibilities is very large, bounded only by the ranges of animals and artifacts that can be combined productively. When the recursive possibilities are considered, the possibilities become infinite (e.g., a camel taking the form of a carriage, with alligators that form themselves into wheels, with starfish that form themselves into spokes, etc.). Children's books are full of many further examples, such as productively combining various human abilities with various kinds of animals (e.g., dinosaurs that talk, tall birds that build sand castles).

As these examples illustrate, the human conceptual ability transcends experience, combining existing knowledge in new ways. Because perceptual symbols are schematic, they can combine creatively to simulate imaginary entities. Wu and Barsalou (1998) demonstrate that when people form novel concepts productively, perceptual simulation plays a central role.

3.1.3. Constraints and emergent properties. Although the productive potential of perceptual symbols is extensive, it is not a simple process whereby any two perceptual symbols can combine to form a whole that is equal to the sum of its parts. Instead, there are constraints on this process, as well as emergent features (Prinz, 1997; Prinz & Barsalou, in press). Presumably, these constraints and emergent features reflect affordances captured through the schematic symbol formation process (2.4.4).

Constraints arise when a schematic perceptual symbol cannot be applied to a simulated entity, because the simulation lacks a critical characteristic. For example, it is difficult to transform a simulated watermelon into a running watermelon, because an entity requires legs in order to run. If a simulated entity does not have legs, then simulating it running is difficult. Interestingly, even if an entity has the wrong kind of legs, such as a chair, it can easily be simulated as running, because it has the requisite spatial parts that enable the transformation.

Emergent properties arise frequently during the productive construction of simulations. As Langacker (1987) observes, combining animate agents productively with running produces emergent properties, with the methods and manners of running varying across humans, birds, horses, crabs, and spiders. Although emergent properties may often reflect perceptual memories of familiar animals running (cf. Freyd, 1987; Parsons, 1987a,b; Reed & Vinson, 1996; Shiffrar & Freyd, 1990, 1993), they may also arise from spatio-temporal properties of the simulation process. For example, when one imagines a chair versus a sofa running, different simulations result. Because people have probably never seen a chair or a sofa run, these different simulations probably reflect the different lengths of the legs on chairs and sofas, the different distances between their legs, and the different volumes above them. Wu and Barsalou (1998) show that emergent properties arise during productive conceptualization as a result of perceptual simulation.¹⁸

3.1.4. Linguistic control of productivity. A foundational principle in Langacker's (1986, 1987, 1991, 1997) theory of language is that grammar corresponds to conceptual structure. One dimension of this correspondence is that the productive nature of grammar corresponds to the productive nature of conceptualization. The productive combination of adjectives, nouns, verbs, and other linguistic elements corresponds to the productive combination of perceptual symbols for properties, entities, processes, and other conceptual elements.

This correspondence provides humans with a powerful ability to control each other's simulations in the absence of the actual referents (Donald, 1991, 1993; Tomasello et al., 1993). Without this productive ability, people could only refer to mutually known referents associated with nonproductive linguistic signs. In contrast, the productive ability to construct simulations through language allows people to induce shared simulations of nonexperienced entities and events. Past events experienced by one person can be conveyed to a hearer who has not experienced them, thereby extending the hearer's realm of knowledge indirectly. Future events can be explored together during planning, decision making, and problem solving, as groups of individuals converge on solutions. Because groups can discuss past and future events, greater teamwork becomes possible, as does more extensive evaluation of possibilities. The productive control of conceptualization through language appears central to defining what is uniquely human.

3.2. Propositions

Another important lesson that we have learned from amodal symbol systems is that a viable theory of knowledge must implement propositions that describe and interpret situations (e.g., Anderson & Bower, 1973; Goodman, 1976; Kintsch, 1974; Norman, Rumelhart, & the LNR Research Group, 1975; Pylyshyn, 1973, 1978, 1981, 1984). A given situation is capable of being construed in an infinite number of ways by an infinite number of propositions. Imagine being in a grocery store. There are limitless ways to describe what is present, including (in amodal form):

As these examples illustrate, different construals of the situation result from selecting different aspects of the situation and representing them in propositions. Because an infinite number of aspects can be propositionalized, selecting the propositions to represent a situation is an act of creativity (Barsalou & Prinz, 1997). Different construals can also result from construing the same aspects of a situation in different ways, as in:

Bringing different concepts to bear on the same aspects of a situation extends the creative construction of propositions further.

Construals of situations can be arbitrarily complex, resulting from the ability to embed propositions hierarchically, as in:

The productive properties of amodal symbols are central to constructing complex propositions.

Not all construals of a situation are true. When a construal fails to describe a situation accurately, it constitutes a false proposition, as in:

Similarly, true and false propositions can be negative, as in the true proposition:

Thus, propositions can construe situations falsely, and they can indicate negative states.

Finally, propositions represent the gist of comprehension. Comprehenders forget the surface forms of sentences rapidly but remember the conceptual gist for a long time (e.g., Sachs, 1967, 1974). Soon after hearing "Marshall gave Rick a watch," listeners would probably be unable to specify whether they had heard this sentence as opposed to "Rick received a watch from Marshall." However, listeners would correctly remember that it was Marshall who gave Rick the watch and not vice versa, because they had stored the proposition:

Thus, propositions capture conceptualizations that can be paraphrased in many ways.

Most basically, propositions involve bringing knowledge to bear on perception, establishing type-token relations between concepts in knowledge and individuals in the perceived world. This requires a conceptual system that can combine types (concepts) productively to form hierarchical structures, and that can then map these structures onto individuals in the world. It is fair to say that this ability is not usually recognized as possible in perceptual theories of knowledge, again because they are widely construed as recording systems. Indeed, this belief is so widespread that the term "propositional" is reserved solely for nonperceptual theories of knowledge. As we shall see, however, if one adopts the core properties of perceptual symbol systems, the important properties of propositions follow naturally. Because perceptual symbol systems have the same potential to implement propositions, they too are propositional systems.¹⁹

3.2.1. Type-token mappings. To see how perceptual symbol systems implement type-token mappings, consider Figure 5a. The large panel with a thick solid border stands for a perceived scene that contains several individual entities. On the far left, the schematic drawing of a jet in a thin solid border stands for the simulator that underlies the concept jet. The other schematic drawing of a jet in a thick dashed border represents a specific simulation that provides a good fit of the perceived jet in the scene. Again, such drawings are theoretical notations that should not be viewed as literal images. The line from the simulator to the simulation stands for producing the simulation from the simulator. The line from the simulation to the perceived individual stands for fusing the simulation with the individual in perception.

The activation of the simulator for jet in Figure 5a results from attending to the leftmost individual in the scene. As visual information is picked up from the individual, it projects in parallel onto simulators in memory. A simulator becomes increasingly active if