Three Major Themes

Three major themes motivate this thesis research. First, I will consider spatial knowledge as a cognitive domain and attempt to characterize the rules that constrain and facilitate learning, guiding what environmental features and kinds of information cotton-top tamarin monkeys (Saguinas oedipus) attend to and use to navigate while foraging. Second, different animal species have unique abilities yet also share universal or common mechanisms. Interspecies comparisons demonstrate that an animal responds to its ecological demands, creating representations that will allow it to maximize its use of the environment. Finally, this thesis is an attempt to clarify the kind of representations possible in a brain without language. As a nonlinguistic species with specific ecological demands, cotton-top tamarins are ideal subjects for studying how these themes interact, showing how a brain without language copes with a complex, three-dimensional environment.

The first major theme of this thesis is the contribution of domain specific as opposed to domain general learning mechanisms to spatial cognition. Accounts of mental structure differ on questions of whether and how brain functions should be categorized and divided, and to what extent these functions are hardwired, innate, or acquired. One way to divide these views is into two categories – domain generality and domain specificity theories. Under this division, Skinner and Piaget fall into the same category of domain general learning theorists (Karmiloff-Smith, 1992). According to Piaget’s (1954) constructivism, the developing child must "construct" their own knowledge of the world using a toolkit of general problem-solving strategies. Children pass through stages of development in which their cognitive structures change in complexity. Piaget’s domain general processes include assimilation and accommodation, which provide a means for a child to deal with new information that may conflict with already stored knowledge, countering this disequilibrium. Behaviorists and associationists similarly posit domain general mechanisms (Fodor, 1983; Karmiloff-Smith, 1992). The child starts out with no built-in knowledge; chaotic sensory input is converted to knowledge through laws of association. Fodor (1983) explains that associationists grant a "set of elements out of which psychological structures are constructed" (p. 27) and associations between the elements formed according to a set of laws. For the associationist, the elements are ideas that are associated; for the behaviorist, the elements are reflexes that are conditioned. These laws are general across domains, and apply to learning in all contexts. The task of these cognitive structures is to produce "sequential redundancies in the mind (or in behavior) which mirror sequential redundancies in the world" (p. 34), picking out correlations in the environment. Fodor (1983) also says that efforts to divide cognition horizontally are further examples of domain general learning theories. In this approach, faculties including "memory, imagination, attention, sensibility, perception" (p. 11), and judgment interact to produce cognition; the mechanisms of each faculty are the same across cognitive domains. So, for instance, the processes underlying memory for musical notes are the same as those underlying lexical or spatial memory.

Fodor (1983) contrasts these horizontal distinctions with vertical ones in which "the intellectual aptitudes (unlike, n.b., the horizontal faculties) are distinguished by reference to their subject matter.… [T]he psychological mechanisms which subserve the one capacity are different, de facto, from those that subserve the other" (p. 15). Fodor’s (1983) proposal is along these lines; he suggests a trichotomous system composed of transducers, input systems, and central processes. Transducers receive environmental stimuli and transmit a signal that preserves the informational content of the stimuli while translating the format so that the next system can process the signal. Central processes receive information from many sources, and consist of high-level cognitive processes such as belief. Between these two levels are the input systems, that do not merely translate information but perform inferences to some extent. He calls these input systems "modules" and says that they are "domain specific, innately specified, hardwired, autonomous, and not assembled" (p. 37). Their operation is mandatory and fast, and associated with fixed neural architecture. Most importantly, modules are informationally encapsulated, in contrast to the central processes. We have only limited access to the information in them, and they are by and large immune to top-down processing.

Many psychologists have expanded on Fodor’s view, but still maintain the "vertical faculty approach." For instance, Karmiloff-Smith (1992) posits the "Representational Redescription" theory, explaining how humans expand on their knowledge through domain specific mechanisms that go beyond Fodorian modules. She disagrees with Fodor’s strict distinction between domain specific modules and domain general central processing, and argues instead that domain specificity may extend its influence to higher cognitive mechanisms. She shifts the focus from tge module to the domain, defined as "the set of representations sustaining a specific area of knowledge: language, number, physics, and so forth" (p. 6), and are not necessarily hardwired, informationally encapsulated, or mandatory. Hirschfeld and Gelman (1994) explain that domains may be characterized as guides to partitioning the world, explanatory frames, and dedicated mechanisms. They are related to modules – distinct, interacting processing systems dedicated to one kind of task. However, modular approaches tend to emphasize "specificity in functional cognitive architecture" while domain specificity theory focuses on "specialization for specific types of knowledge" (p. 10). Development, when viewed as domain specific, entails phases rather than Piagetian stages; domains and microdomains evolve and progress at different rates, though they may each go through similar transitions. In contrast to Fodor’s nativism suggesting that modules are genetically prespecified, most domain specificity theories focus more on innately specified predispositions that guide modularization. Overall, theories that have expanded on Fodor’s modularity hypothesis appear to loosen his tight definitions and expand vertical divisions to higher cognition.

Support for domain specific learning over domain general processes come from a variety of sources. Fodor (1983) points to the fact that human adults have differential abilities; for example, someone with a good memory for numbers may not be good at remembering the location of landmarks. From the developmental perspective, children do not appear to progress in general senses, as in Piaget’s stage model; instead, they appear to make progress in some areas before others that are best broken down by domain (Karmiloff-Smith, 1992). Also, neurological lesion patients often show specific, characteristic breakdown patterns. People suffering from agnosias and aphasias often present with specific deficits in one domain but not another, such as face recognition but not object recognition. Fodor (1983) also points to instances in which people process the same type of information differently depending on the domain. For example, auditory nonspeech is analyzed differently than speech sounds are; phonemes isolated from an utterance are heard as "whistles" and "glides" as opposed to the onset of a consonant as perceived when part of a word. Furthermore, we often lack conscious awareness and access to many "modular" processes; visual illusions and the phoneme restoration effect are two examples of mandatory mechanisms that are not influenced by top-down knowledge (Fodor, 1983). Finally, the claim that modules and domains operate and develop according to innate specification is supported by the speed and nature of knowledge acquisition in particular domains. For instance, Chomsky (1965) proposed a language organ and a universal grammar to account for the shortcomings of the behavioral view of language acquisition. The poverty of stimulus argument holds that language cannot be learned merely through laws of association and reinforcement. The child learns so fast that it seems unlikely that she is provided with enough evidence to formulate all the rules in such a short time frame.

Cosmides and Tooby (1994) explain that natural selection has brought about domain specific mechanisms instead of domain general processes because of the former’s "speed, reliability, and efficiency" (p. 89). "Adaptive problems" require domain specific mechanisms because they "require different solutions, and different solutions can, in most cases, be implemented only by different, functionally distinct mechanisms" (p. 89). When an animal confronts a problem in its natural environment, the cues it uses and the problem-solving mechanisms it falls back on will depend on the nature of the problem itself. The trouble with a domain general learning mechanism is that it requires that the animal consider all possibilities in any given circumstance and learn about completely different concepts with the same processes. When trying to identify objects, for instance, Fodor (1983) explains:

Modules and domains help to limit the amount of information the animal must consider, and the number of potential hypotheses to explain an event. As Karmiloff-Smith (1992) puts it: "Domain specific constraints potentiate learning by limiting the hypothesis space entertained" (p. 11). They also provide a relatively automatic guide for what mechanisms are most relevant in the particular context. For instance, the machinery required for a human child to learn language may be very different than that needed for face recognition.

By considering spatial knowledge as a cognitive domain, we subject it to constraints and dedicated mechanisms that guide spatial learning and set this form of cognition apart from other domains. The distinction between domain specific and general processes manifests in spatial research in the competing views of those who endorse a mapping system that draws on specific kinds of representations, and those who support an asssociationist view by which spatial knowledge is acquired through the same processes involved in all forms of learning. Tolman (1988) discusses the debate between supporters of the view that rats learn mazes by a series of conditioned stimulus-response associations, and the "field theorists" who propose a cognitive map. Tolman’s (1988) map encodes "routes and paths and environmental relationships" (pp. 440-441), and constrains what kinds of information the rat attends to. Without a kind of mapping system, he argues, the rat would be unable to extract relevant cues and ignore extraneous stimuli for navigation tasks. What Tolman is describing is a learning system that is constrained by a set of rules that guide what features an animal will pick out of the environment, how it will parse its perceptual experience to make sense of it, and how it categorizes objects and events.

Landmarks are spatial objects, the elements of a "map" encoded as place representations. The set of rules characteristic of the spatial domain will determine what environmental and object properties draw an animal’s attention. In this thesis, we investigate how domain specific learning contributes to the way animals treat spatial objects – specifically, what landmark features tamarins consider most important. A great deal is already known about tamarins’ object knowledge. For instance, they appear to expect objects to behave according to the laws of gravity (Hood, Hauser, Anderson, & Santos, 1999), and they also seem to count objects in space. In the domain of food, rhesus monkeys prefer color properties to the shape of objects in determining what is edible (Santos, Hauser, & Spelke, in submission). On the other hand, in the domain of tools and artifacts, cotton-top tamarins preferentially notice shape and size changes, ignoring color and texture (Hauser, 1997). As Hauser writes, "This makes sense since color and texture do not affect the tool’s function, whereas shape and size do" (p. 285). Investigating how tamarins respond to objects that serve as landmarks promises to expand on this previous research and demonstrate how a spatial setting influences object knowledge.

Studying cotton-top tamarins’ spatial ability may also provide the opportunity to gain a better understanding of spatial knowledge in other animals using an interspecies comparisons approach, another theme of this work. Many principles of navigation are universal or widespread in animals. For instance, processes such as dead reckoning and piloting are remarkably similar in insects, birds, and mammals (Gallistel, 1990). However, species-specific traits reinforce the idea that different ecological constraints generate differing innate predispositions (Hauser, 2000). Exploring the principles that guide foraging in tamarins may help both to clarify the principles at work in all navigating animals and to illustrate how ecology influences species differences.

Cotton-top tamarins are an arboreal species, producing a demand for three-dimensional spatial encoding. In the wild, these monkeys consume fruits of different kinds of trees and plants, demanding that the monkeys be able to differentiate between vegetative species (Garber, 1989). Also, as individual members of plant species are often widely separated in space, the tamarins must be able to encode the location of multiple food sources in order to forage more efficiently. Garber (1989) studied moustached tamarins (Saguinas mystax) and saddleback tamarins’ (Saguinas fuscicollis) foraging patterns in the Amazon Basin of northeastern Peru, and found that the monkeys fed on fruit from more than 20 tree species. Each day, the groups focused on a select number of trees from a target species, usually going directly from one tree to the next tree of the same species. Based on these observations, Garber speculated that the foraging is goal-directed, and that the tamarins encode the locations of a large number of feeding sites such that they can minimize random foraging movements. In a subsequent paper, Garber and Hannon (1993) compared these foraging patterns with a computer-generated model, and concluded that "although tamarins may rely on olfactory cues to locate nearby feeding sites, their foraging patterns are better explained by an ability to maintain a detailed spatial map of the location and distribution of hundreds of feeding trees in their home range" (p. 827). Controlled experiments using spatial tasks will elucidate how these ecological demands may have produced predispositions in the tamarins differing from those of other species.

The evolutionary perspective offers comparative power in the sense that we can test claims about "uniquely human" abilities and precursors found in nonhuman animals as homologous traits, while taking care to distinguish these from homoplasies. Whereas a homology is a trait shared by two species due to inheritance from a common ancestor, homoplasy refers to interspecies resemblance due to parallel evolution or convergence, without common ancestry. Homoplasies involve adaptive solutions arrived at independently by two species, and speak to how similar ecological constraints yield particular responses. However, different mechanisms may underlie homoplastic behaviors. Homologies, on the other hand, suggest mechanistic similarity, such that describing traits in one species helps in the understanding of related species behavior.

Preuss (1995) explains that "acknowledging a diversity of minds … suggests new questions about the human brain and cognition" including: "Why this outcome and not some other? What specific cognitive capacities were selected for in human evolution? How were the components of neural and cognitive systems present in our primate ancestors modified to produce new systems in humans?" (p. 1239). Some psychologists claim that certain kinds of knowledge are available only to humans. For example, Karmiloff-Smith’s (1992) theory of "Representational Redescription," whereby domain specific representations are re-represented repeatedly to become more available to other domains, precludes nonhuman animals. In her theory, information is first encoded procedurally, then is re-represented at increasingly abstract and productive levels. Even at these higher levels, however, these representations may not be consciously accessible or subject to verbal report. Even so, redescription is essentially a process of "explicitation" whereby implicit knowledge becomes increasingly explicit, and is undiscovered in other animals. According to this theory, only humans can treat their own representations as objects and manipulate them to generate further knowledge. She writes: "Intra-domain and inter-domain representational relations are the hallmark of a flexible and creative cognitive system. The pervasiveness of representational redescription in human cognition is, I maintain, what makes human cognition specifically human" (p. 192). By investigating the representations of tamarins, we can test the merits of ideas like this, and ask why there is a distinction between humans and all other animals and if animals demonstrate behaviors that are fundamental to but not sufficient for these abilities.

Drawing a distinction between humans and nonhuman animals relates to the third theme – the study of representation without language. As a nonlinguistic primate species, tamarins’ navigational abilities may be complex but without the benefit of linguistic encoding. As Thinus-Blanc (1995) put it, "this approach allows us to have an insight into ‘pure’ spatial mechanisms without the interference of highly symbolic tools mediated by the use of language in human beings" (p. 241). Some psychologists claim that many kinds of representations and knowledge are not available to an animal without language. For instance, in its strongest form, the Whorfian hypothesis argues for linguistic determination of thought, whereby language directly effects the way that people think about and see the world. Whorf (1956) explains:

Whorf hypothesized that linguistic patterns determine what the individual perceives in the world, and that people who speak different languages will thus have fundamentally different modes of thinking and world-views.

Related to the Whorfian Hypothesis is Gopnik and Meltzoff’s (1986) "Specificity Hypothesis" concerning the relationship between linguistic and cognitive abilities, whereby the acquisition of words relates to the development of concepts. For instance, words about "disappearance", and "success" and "failure" appear related to the child’s emerging understanding of objects and of means-end relationships during the one-word phase, respectively, around 12 months. In a cross-cultural study of spatial language, particularly prepositions, Pederson et al. (1998) found that "a community’s use of linguistic coding reliably correlates with the way the individual conceptualizes and memorizes spatial distinctions for nonlinguistic purposes" (p. 557). One claim is that the language available to the person constrains the type of thoughts the person can have. To counter the inference that animals completely lacking human-like language would be at a total loss when it comes to cognitive abilities, one could flip the problem around and claim that it is cognition that constrains linguistic ability. However, the correlation between linguistic and cognitive abilities remains, suggesting that a nonlinguistic animal may not be capable of high-level, complex ideation. This hypothesis calls into question the ability of nonhuman animals to create abstract representations because they lack linguistic labels. Showing that animals can form and utilize concepts will help challenge strong claims that animals need language to understand abstract relationships.

Spatial representations are integrative, combining information from multiple modalities and modules. Cheng (1986) and Hermer and Spelke (1996) have found that rats and human infants cannot conjoin geometric and nongeometric information to navigate after disorientation; they are unable to reestablish their position by combining two kinds of information, and instead use only the global shape of the environment (see Chapter 3 for further discussion). Hermer-Vazquez, Spelke, and Katsnelson (1999) suggest that these organisms fail because they do not have the necessary linguistic machinery. Human children start to solve the task when they begin to use spatial language allowing for the integration of geometric and nongeometric information (Hermer, 1994). Furthermore, adults who are verbally shadowing (repeating verbal material syllable by syllable as it is spoken without pausing for more than 2 s) perform like rats and human infants after disorientation, unable to combine knowledge from different modules. Verbal shadowing may occupy the language faculties, rendering adults incapable of using language to combine the information. Hermer-Vazquez et al. (1999) suggest that the language faculty is a "uniquely human combinatorial system of representation" (p. 34) that "plays a role in the generation of the flexible spatial performance that is unique to humans among extant species" (p. 32). They claim that two characteristic of language enhance spatial representations: (a) a lexicon consisting of terms referring to items from different domains (such as objects, colors, and spatial properties); and, (b) grammar allowing these terms to be combined in prepositional structures (such as "east of the blue building"). A study of nonlinguistic nonhuman primates provides an opportunity to explore this possibility: as tamarins do not have access to language, they should be constrained in their ability to integrate across modules to solve complex navigational tasks flexibly.

Spatial Representations

Gallistel (1990) says: "The brain is said to represent an aspect of the environment when there is a functioning isomorphism between some aspect of the environment and a brain process that adapts the animal’s behavior to it" (p. 15). He goes on to explain that functional isomorphisms are one-to-one correspondences between "distinct systems of mathematical study" resulting in a "parallelism of form" such that it is "exploit[ed] to solve problems in one domain using operations belonging to the other" (p. 15). In particular, spatial representations consist of "representing entities" in the brain, such as Cartesian coordinates, numbers, and vectors, that correspond to represented entities in the environment, such as positions on a plane, distances, and angles. Animals use representations to solve problems in the represented environment.

It is difficult to know exactly what kinds of representations humans possess, no less other animals. Many problems can be solved with different kinds of representations. One distinction is between a snapshot system and more abstract, relational concepts. Analogous to Tolman’s distinction between "field theory" and behaviorist hypotheses of spatial learning, Menzel et al. (1996) describe the differences between the cognitive map concept and the local rule concept. Navigation by the local rule is accomplished through many separate perceptual memories. Several models of spatial navigation posit that animals encode a series of snapshots of the environment, and move in ways that minimize discrepancies between current and remembered visual perceptions (Wehner, 1996). Wehner, Michel, and Antonsen (1996), studying navigation in social hymenopterans, conclude that "there is not need to invoke the concept of the mental analogue of a topographic map – a metric map – assembled by the insect navigator. The flexible use of vectors, snapshots, and landmark-based routes suffices to interpret the insect’s behavior" (p. 129). In these studies, desert ants (genus Cataglyphis) and honeybees (Apis mellifera) respond to distorted arrays of landmarks in a manner indicating the use of a two-dimensional visual snapshot. The snapshot is retinotopically fixed, so that an insect may only match a stored with a current representation by moving into the position in which the original image was encoded. Poucet (1993), in a review of the literature, comments that this kind of representation cannot account for some of the complexities of animal behavior, as "it emphasizes the associationist view of spatial behavior, which does not fit the flexibility of observed map-based behaviors" (p. 165). Poucet argues instead for a model based on place representations abstracted from continuous perceptual experience, a representation that preserves only the relevant information in navigation, and he points to a range of flexible behaviors such as novel route formation.

Even if one were to invoke a "map," there is still the problem of distinguishing between different kinds of maps. A "route map" is a series of memory images stored during previous trips down a route from one place to another (Gallistel, 1990). In one sense, this is another version of the "snapshot" model, in that it is simply a sequence of stored perceptions. The animal uses the sequence when travelling down that route again. A "vector map," in contrast, is a host of stored locations and vectors, allowing for the animal to calculate routes by vector addition and subtraction. Menzel et al. (1996) explain that vector integration may be considered as an extension of dead reckoning, the process by which animals continually update their position relative to an origin by integrating their velocity (directed speed) with respect to time. Dead reckoning, or path integration, yields vectors that specify distance and direction. Animals could form vector maps by integrating a series of separate homing vectors. Through vector addition and subtraction, new vectors are calculated between various geostable locations, constituting a "vector map." Dyer (1996) describes the vector map as a "metric map" that references an allocentric coordinate system, possibly in a Cartesian format, but that is constructed from recorded egocentric, retinal images. Menzel et al. (1996) observe that bees demonstrate map-based behavior, selecting novel routes when they see local cues within a range of two meters, and at certain times when celestial cues are unavailable. They claim that the evidence supports the vector map concept.

One major problem is distinguishing between a vector map and a cognitive map, as many cognitive map-like behaviors may be explained with a vector map. Complicating this issue is the difficulty in defining the cognitive map precisely, such that it is often referred to in the sense of what other types of representations cannot do. If one wants to ask whether an animal has a cognitive map, the question becomes more a problem of assembling a mountain of evidence that would be adequate to prove that another form of representation is not enough. Bennett (1996) argues that the cognitive map is not a useful hypothesis, and that no experiment to date has undeniably proved the existence of a cognitive map even in human beings. However, the conceptual issues underlying the term "cognitive map" are useful in the study of animal navigation. Gallistel (1990) describes a cognitive map as isomorphic in the sense that stimuli in the environment correspond in a one-to-one relationship to their representations in the navigator’s representation, and as multimodal, thus formed through the integration of information taken from the environment in many ways (e.g., visually, haptically) over multiple intervals of time. Cognitive maps are flexible, allowing for generalizations and novel solutions to navigational problems and for communicative behavior indicative of abstract representations.

The flexibility of the cognitive map arises because it is formed by integrating both geocentric and egocentric representations, such that features of the environment are represented and considered during navigation even if they are out of the immediate field of the animal’s perception. Gallistel and Cramer (1996) explain that the map is constructed by integrating geocentric information acquired while dead reckoning with egocentric information gathered while exploring the environment and observing landmarks and terrain features. Dead reckoning allows for a coordinate framework in which the origin is the animal’s starting point; vectors between the animal and landmarks are computed by perceptual systems, and in this way the terrain features are given "addresses" on the coordinate system. Using this kind of representation, the animal can calculate its position by recognizing landmarks and can plan efficient routes by taking into account places not within its immediate perception.

Communicative behavior in animals shows that they encode the location of distant places and can formulate a route to return there. For example, to recruit other hivemates to exploit a food source, honeybees (Apis mellifera) use "dance language" that transmits information about the quality and location of the source. Round dances communicate nearby locations, while the "waggle dance" pertains to distant food sources. During the waggle dance, the communicating bee moves in the form of a figure eight; the dance transmits information about the angle of the goal location relative to the sun’s current azimuth (Gallistel, 1990), while the number of turns of the dance is correlated negatively with the absolute distance (Hauser, 1996). In one experiment, bees who found food in an "artificial" location in the middle of a lake could not recruit other bees with the dance (Gould & Gould, 1988). Hypothetically, the hivemates must have maintained a map representing a region that could not possibly contain a food source, and by referencing that representation, they decided the communicating bee was mistaken. Menzel and Halperin (1975) studied purposive behavior in chimpanzees (Pan troglodytes) approaching hidden, distant goals that were either toys or food. One leader chimp moved toward the goals in the presence of several other chimps. The leader’s locomotion served as objective communication to the others. The rate at which the leader approached the goal transmitted information about the identity of the hidden object, influencing whether the others followed to the goal, and the response of the other chimps in turn influenced whether the leader reached the goal. This behavior indicates a representation encoding the identity of hidden objects, allowing the chimps to plan routes and formulate strategies to retrieve goals in social settings. The communicative behavior of honeybees and chimps demonstrates an understanding of the existence of places outside the field of perception

Tolman (1988) suggests that one crucial sign that an animal has encoded a map of its environment is its ability to generate novel paths they have not traveled before, but are computed from stored representations. His rats were able to choose the correct arm of a radial maze to find a goal location they learned previously by traveling through a square maze. Tolman claims they must have abstracted a map-like representation over many intervals independent of the temporal sequence of location-learning. Thinus-Blanc (1995) describes mental maps as "plastic" or "flexible" in that they are characterized by a "reorganization of spatial information independent of the sequence during which it has been acquired.… The orientation process relies not on rigid routes, but on the precise localization of the places" (p. 262). Animals that have representations integrated from a series of perceptual experiences are thus able to generalize, forming novel routes that minimize distance traveled.

Experiments studying novel route formation in primates test their ability to solve the "traveling salesman problem," which may be informally defined as finding the shortest route for a "salesman" to visit all the cities on a list at the lowest cost. Menzel (1973) carried chimpanzees around an outdoor field, showing them the hiding places of up to 18 food stores. When later released into the space, the apes appeared to generate a retrieval route, different from the route during training, which minimized the total distance traveled. This finding may indicate that the chimpanzees had created a metric map recording the geometric relations between the food sites, and then had derived the most efficient novel route to retrieve the items from this representation. However, this is confounded by the fact that the animals could have simply chosen the site within their field of vision that was closest to their current location, avoiding the work of using map. Savage-Rumbaugh, McDonald, Sevcik, Hopkins, and Rupert (1986) showed that bonobo chimpanzees (Pan paniscus) who learned the location of 17 different feeding stations in a heavily wooded area could remember these locations and, when tested four months later, could lead a human experimenter to a location displayed by a photograph. They appeared to choose paths minimized for distance. The target locations were out of sight from the initial starting point, indicating that the apes may have created a cognitive map. However, it is unclear whether the routes were really novel, or if the chimps had traveled over them before.

One explicit test of primates’ ability to solve the "traveling salesman problem" was conducted on vervet monkeys (Cercopitheus aethiops) by Cramer, and is reported in Gallistel and Cramer (1996). Like Menzel (1973), Cramer carried the vervet monkeys around an enclosure, showing them the locations of hidden food. In contrast to Menzel’s chimps who could learn up to 18 locations, Cramer’s vervets could only remember the locations of 6 or fewer stores of food. However, Cramer clearly showed that the vervets used a strategy to find the optimal route minimizing the total distance traveled, looking ahead three steps (taking into account the next three destinations), instead of using a simple solution such as the nearest-neighbor algorithm which calls for looking only one step ahead. In order to calculate the optimal route, the vervets must have in mind multiple destinations on a coordinate map system, such that they can compute vectors and routes to formulate an overall plan. These monkeys were not simply acting on immediate perceptual information, but instead were tapping complex representations.

If an animal were to maintain a cognitive map representing the overall shape and landmark configuration of an environmental space, or a representation that isomorphically encodes a relationship between two or more landmarks, then the encoded relationship should be flexible and subject to generalizations. Abstract spatial relationships allow animals to understand a fixed relationship between two or more moveable landmarks and to generalize these spatial rules to novel situations. One way of getting at this kind of conceptual encoding is a paradigm that keeps landmark configuration constant while varying the overall scale. If an animal encodes the location of a food source by its relation to a group of landmarks, then expanding or contracting the size of the configuration should not alter a general rule such as "search midway between the two trees." If, on the other hand, the animal encodes the target location in a more basic, associative way, then a rule such as "search 10 ft from that specific tree" would be more likely. This recalls the distinction made by Tolman (1988) between associative spatial learning and more symbolic and cognitive processes. Clark’s nutcracker (Nucifraga columbiana) has been shown to apply the rule "midway between two landmarks" (Kamil & Jones, 1997). However, in another experiment on pigeons (Columba livia) and humans (Spetch et al., 1997), while humans were able to locate a target in the center of four landmarks when the configuration was contracted or expanded, the pigeons were unable to do so and instead resorted to searching at an absolute distance from one of the four landmarks (see Chapter 5 for a more complete description and comparison of these studies). Perhaps the need to remember thousands of locations of caches has required the nutcracker to use more complex spatial representations than the pigeon uses. This dichotomy points to possible evolutionary or ecological sources of the differential abilities, as the two avian species occupy different niches.

The above discussion indicates that symbolic spatial encoding is characterized by atemporal and isomorphic representations, the generation of novel paths, and the ability to generalize spatial relationships to altered environments. Spatial learning is not simple associative learning by conditioning and pairing. Instead, this domain of cognition is constrained in certain characteristic ways. Thus landmarks, if considered as objects in an environment, are attended to differently than objects perceived in other domains. In a natural environment, animals must use only relevant information for the task. In the domain of landmark-based navigation, geometric and positional properties may generally be considered more reliable than nongeometric, as in the natural environment the latter is at greater risk of change than the former. Experiments on rats (Rattus norvegicus) by Biegler and Morris (1996) indicate that the featural properties of landmarks are efficiently used to locate a food cache only if during training the landmark was maintained at a constant geometric position relative to the arena. Geometric properties may be the most salient features to navigating animals. One study by Cheng (1986) revealed that rats could use nongeometric information in a rectangular box to differentiate between the corners of an otherwise ambiguous environment only when it did not conflict with the geometric information. When these properties are conflicted, rats use geometric information exclusively. Hermer and Spelke (1996) similarly found that disoriented toddlers fail to use featural cues to determine their heading, but do use those cues when they are oriented See Chapter 3 for a more complete discussion of these studies). Disorientation appears to render nongeometric information unreliable for toddlers and rats. Similarly, Greene and Cook (1997) found that rats can still use a configuration of landmarks when their visual cues are altered. However, if the objects are arranged in a novel configuration, the rats fail to use individual landmark identities to search for food. These studies show that the processes underlying spatial learning are constrained by the demands of this domain.

Nonhuman primates also seem to pay more attention to the geometric than nongeometric information of landmarks. For instance, in one study by Tinklepaugh (1932), chimpanzees who witness an experimenter hide food in one of two containers in eight pairs arranged in a circle often search in the geometrically correct location after the entire circle is rotated. Menzel and Menzel (1979), using a habituation procedure, showed that captive marmosets (Saguinas fuscicollis) are not only able to recognize changes in the objects constituting a familiar environment, but that they are most surprised by landmark exchanges, next by positional changes, and least by orientation changes. When position is held constant, the primates are able to detect the change in object identity. Positional change is probably more salient than orientation change because one fundamental property of landmarks is their geometric stability. That the marmosets are least surprised by orientation changes insinuates that the monkeys recognize that it is the same object, but changed in a less interesting and more understandable way. Furthermore, the marmosets dishabituated and responded to a greater extent when an object was changed from a vertical to a horizontal orientation between familiarization and test trials than for the inverse transformation. This result contrasts increased looking time for vertically oriented objects displayed by free-ranging, terrestrial rhesus macaques (Macaca mulatta). Menzel and Menzel (1979) suggest that this makes a lot of ecological sense: these monkeys may be most surprised by changes from the familiar to novel. Arboreal primate species may be more accustomed to verticality, whereas terrestial species may be more familiar with horizontally oriented objects. Again, ecological demands produce differences in object knowledge and spatial representations.

Constrained learning is also evident in research done in the field. Garber, Paciulli, and Dolins (Garber & Dolins, 1996; Garber & Paciulli, 1997) have investigated foraging strategies in moustached tamarins and capuchin monkeys (Cebus capucinus). In both sets of studies, there were 13 to 17 platforms at which either real or sham bananas were placed. Both primate groups most readily respond to platforms that consistently contain food, while olfactory and visual cues play lesser roles in platform-exploration choice. Also, in one condition, a goal-sign (either a yellow block or a red flag) was always placed close to platforms with real bananas. Tamarins appear unable to use this information, performing at chance levels. Some capuchins did show some indication of goal-sign usage, however, and Garber and Paciulli (1997) propose an ecological explanation, speculating that because tamarins rarely exploit concealed food sources (whereas capuchins do), they may not need to encode visual features of nearby landmarks, instead relying on visual sighting of food and distant landmarks. However, perhaps a simpler explanation can be derived from Biegler and Morris’s (1996) studies described above, in which rats did not use featural information of landmarks unless the landmarks were positionally constant during the first trials. Tamarins may not be able to make the simple association between the red flag and the food source because the position of the red flag was not held constant. Because of their varying positions, the tamarins may not have perceived the red flags as landmarks, and thus did not consider them as reliable indicators of food location.

Cheng (1986) and Hermer and Spelke (1996) hypothesized that different kinds of information are encapsulated in the brain, subject to dedicated processing mechanisms and free from cross-talk. Rats and toddlers appear to ignore featural information though control trials reveal that they do notice and can recall the information. The fact that these subjects fall back almost exclusively on geometric properties to reestablish heading upon disorientation suggests that geometric and featural information are stored and processed in distinct "Fodorian" modules – the "geometric module" and the "nongeometric module." Similar to the way the visual system is comprised of separate modules individually processing such information as color, direction, and movement, spatial information may be modularized at low levels and integrated with greater complexity as it is passed on to higher levels. In the visual system, these modules appear to recruit select populations of cells, organized into columnar neural structures.

In a similar way, modularization of spatial information processing may be manifest in neural architecture. The hippocampus appears to be directly involved with spatial cognition. For instance, Sherry and Duff (1996) have found that birds who store caches of food tend to have larger hippocampi than birds who do not, and that the size of this brain region may vary seasonally due to greater dependence on spatial reorientation during certain parts of the year. Hippocampal lesions disrupt cache retrieval in birds, and also disturb rats’ performance in spatial task such as the Morris water maze (for review, see Gerlai & Clayton, 1999). Some groups of hippocampal cells in CA3 fire selectively when the animal is in specific locations within an environment, and thus have been called "place cells" (McNaughton, Knierim, & Wilson, 1996). These cells are controlled by visible landmarks, and may or may not be dependent on directional information. A sequence of these cells firing allows for "place fields" that rotate if a landmark array is rotated. The parietal and posterior cingulate cortex and the subicular complex contain head-direction cells that fire for specific bearings relative to a landmark (McNaughton et al., 1996). O’Keefe and Burgess (1996) found that some rat hippocampal cells respond to absolute distance from environmental features, while others respond to proportional distances between two walls. Different forms of spatial knowledge – location, direction, and distance, for instance – may be evaluated separately in the brain.

Some researchers have tried to define the conditions under which animals can integrate modularized information. Hermer-Vazquez et al. (1999) suggest that in humans, information from the geometric and nongeometric modules are integrated only once the person has access to language. Spatial language allows for prepositional structures conjoining referents from different domains. For instance, the statement, "to the left of the blue wall" puts together directional ("left of") and featural ("blue wall") aspects to specify one corner of a rectangular room uniquely. Similarly, Arterberry (1993) claims that infants’ ability to perceive object properties over time depends on spatiotemporal integration that arises around 12 months of age. Xu (1999) suggests that linguistic labels allow human children to integrate information from the "what" and "where" pathways, and that language thus plays a crucial role in humans’ ability to individuate objects.

These theories imply that nonhuman animals may have trouble integrating information processed separately. However, despite the absence of language, many animals are capable of abstract spatial thinking. Naturalistic experiments demonstrate that some primates classify both landmark and food types and use these categories to forage efficiently. This kind of cognitive ability is essential for species that forage in areas of tremendous ecological diversity (Menzel, 1997). Menzel (1991) reports that Japanese macaques (Macaca fuscata) search differently when they find distinct kinds of food items. If confronted with an akebi fruit on the ground, the macaques look up, as though they were expecting to find more food overhead in the trees. Furthermore, the monkeys look only at akebi plants and not at non-akebi plants. However, if confronted with chocolate on the ground, the monkeys confine their search to the ground, as though expecting that more chocolate can be found there and not in the trees. These studies suggest that the macaques could recognize food type, and made use of this information in constructing a search strategy. Similar results were found with long-tailed macaques (Macaca fascicularis) in an experiment investigating their ability to use landmark kind information (Menzel, 1996). When released into an environment with a series of landmarks that may be divided into structural classes, the monkeys search near landmarks that are similar to the landmark by which an obviously visible cache of food is found. Menzel proposes that the ability to track landmark kinds is crucial to the monkeys’ survival, as fruits ripen at intervals that are often synchronous within species. One important caveat to note, however, is that upon reintroduction into the environment, the monkeys often place more weight on the position at which the food was found in the first exposure, reemphasizing the necessity for landmark-stability in spatial navigation.

It is apparent that nonlinguistic species do have a variety of spatial mechanisms at their disposal. However, even in cases where animals appear to be using abstract concepts such as those we attribute to ourselves, it is likely that the mechanisms underlying those behaviors may not be identical. What does seem constant across species is the subdivision of spatial mechanisms into compartmentalized information-processing units dedicated to specific calculations, and the subsequent need to weight the various sources of information. The process of weighting and integration may require completely ignoring certain cues, or compromising when contradictory stimuli are detected. Animals possessing basic processes such as binocular disparity, retinal image size, looming, motion parallax, and other mechanisms to judge angles and distances, and capable of encoding information about the individual identity of both global and proximal cues and their "addresses" within an overall configuration, must choose what information is most relevant and reliable in the given situation. The kind of spatial representation the animal encodes depends on both the information that gets into the system and the information that gets past this filter.

The Questions and the Plan of the Thesis

The representation of geometry as a configuration of landmarks and the ability to have abstract concepts may serve as the basis for an animal’s cognitive map. We will not be able to demonstrate absolutely the existence of a cognitive map in tamarins, at least within the space of this thesis, but we can attempt to characterize the nature of the tamarins’ spatial representations. We hope to answer such questions as: What kinds of representations are possible in a brain without language? Can tamarin monkeys integrate information processed in different modules without language? Can monkeys represent an abstract geometric configuration? Would the concept be irrespective of scale, temporal aspects of acquisition, the particular learning condition, and the particular location? What kinds of information get into the monkey’s representation? What constraints are put on tamarins’ spatial learning, as a domain of knowledge? Does the domain constrain what landmark features are important during navigation? And finally, how do ecological demands influence tamarin spatial abilities? Does the fact that tamarins are an arboreal species bear on the three-dimensionality of their spatial cognition? What sorts of strategies for efficient foraging are available to a species dependent on a variety of tree species spread over large ranges in the wild?

The second chapter presents the general procedures used in the experiments, and a brief natural history of the tamarins that provides a background for questions of how ecological demands influence behavior. The first experiment, presented in Chapter 3, is a replication of Cheng’s (1986) and Hermer and Spelke’s (1996) studies in which subjects reoriented in a rectangular enclosure with and without featural cues. The tamarins will provide a comparative case against the rats and the humans. Through this paradigm, we can ask whether tamarins can solve the task with no training, thus integrating geometric and nongeometric information without language. The second experiment, in Chapter 4, builds on the first. In the rectangular room, the monkeys must use the rule "to the right of a cue" or "to the left of a cue," when the cue is a striped wall. For the second study, tamarins learn to search "above a cue" or "below a cue," when the cue is a small colored cylindrical object. The task tests the tamarins’ ability to form an abstract concept that integrates information from two distinct modules. Furthermore, once the subjects have learned the rule, we can test which landmark properties the monkeys deem most important. By altering the features of the landmark itself and comparing the monkey’s response to the altered landmark to the response to the familiar landmark, we can judge how salient a feature is by the extent of the differences in response. This work explores the idea of spatial knowledge as a cognitive domain constrained by rules that dictate what features animals pick out of their environment, and attempts to define what object properties are most important in this domain. The third experiment, presented in Chapter 5, focuses on tamarins’ ability to form and use abstract concepts. In this experiment, monkeys learn to search midway between two points given limited variability during training. Upon test, we manipulate this configuration to see the extent to which the monkey can generalize the rule, and if they encode the position of the goal as a point in a configuration or as a point calculated by single landmark bearings. We test the monkeys on expansions, rotations, and translations to examine the limits of their generalizations. Furthermore, some rotations are made in the vertical plane, to test whether these arboreal monkeys’ representations can be manipulated through three-dimensional space. The final section of the thesis, Chapter 6, is the General Conclusion, and is intended to tie together the ideas of the integration of modular information, nonlinguistic representation, abstract concepts, domain specificity, and requisites for a cognitive map. This chapter offers speculations about the significance of these experiments, and suggestions for future avenues of investigation.