Infants' ability to use object kind information for object individuation

Abstract

The present studies investigate infants reliance on object kind information in solving the problem of object individuation. Two experiments explored whether adults, 10- and 12-month-old infants could use their knowledge of ducks and cars to individuate an ambiguous array consisting of a toy duck perched on a toy car into two objects. A third experiment investigated whether 10-month-old infants could use their knowledge of cups and shoes to individuate an array consisting of a cup perched on a shoe into two objects. Ten-month-old infants failed to use object kind information alone to resolve the ambiguity with both pairs of objects. In contrast, infants this age succeeded in using spatiotemporal information to segment the array into two objects, i.e. they succeeded if shown that the duck moved independently relative to the car, or the cup relative to the shoe. Twelve-month-old infants, as well as adults, succeeded at object individuation on the basis of object kind information alone. These findings shed light on the developmental course of object individuation and provide converging evidence for the Object-first Hypothesis [Xu, F., Carey, S., 1996; Xu, F., 1997b]. Early on, infants may represent only one concept that provides criteria for individuation, namely physical object; kind concepts such as duck, car, cup, and shoe may be acquired later in the first year of life.

Introduction

Perceiving the world in terms of distinct three-dimensional objects is one of our most fundamental cognitive capacities. The study of object perception may be divided into two enterprises: explaining object recognition (i.e. recognizing a given object as a member of a known kind) and explaining object individuation (i.e. representing a given visual scene in terms of distinct objects). In addition, the question of the relations between object recognition and object individuation arises. For example, according to Marr (1982), prior to both individuation and recognition the visual system constructs a representation of continuous surfaces. He speculated that it is only after we recognize objects of particular kinds that we are able to parse the layout into individual objects. Similarly, there is a philosophical tradition which holds that it is in principle impossible to individuate objects without knowing what kinds the objects belong to (Geach, 1957; Gupta, 1980; Wiggins, 1980; Hirsch, 1992). Our knowledge of the kinds of things there are in the world certainly helps us establish representations of the distinct objects in a visual scene; our knowledge of ducks and cars, for example, leads to a representation of two objects rather than one when a toy duck sits on top of a toy car with no clear boundary between them.

Although kind information is most definitely useful in object individuation, it is not always necessary. Several other types of knowledge play a role, the most fundamental of which is spatiotemporal information. Physical objects are bounded, coherent wholes which maintain their integrity as they move through space on spatiotemporally continuous paths (Spelke, 1990). Therefore, the most unambiguous evidence that an array contains two distinct objects is provided when they are simultaneously visible and separated in space, and indeed very young infants use such information to establish representations of distinct objects (Baillargeon and Graber, 1987; see also Xu and Carey, 1996for a demonstration with older infants). Problems of object individuation arise under conditions where perceptual access to boundaries or to spatiotemporal continuity is lost. Two conditions that generate ambiguity are of particular interest to the present discussion: ambiguity arises in cases of shared boundaries, which is often called the problem of segregation or segmentation (Needham et al., 1997). Upon seeing a complex display such as what is shown in Fig. 1, one might ask: is that duck-on-a-car one object or two? Ambiguity also arises in cases of successive appearances of objects. When an object disappears from view followed by an object reappearing, the question of object individuation arises: is there one object or two involved in this event?

Under such ambiguous conditions, very young infants have been shown to exploit spatiotemporal information, if available, to individuate objects. Kellman and Spelke (1983)showed that infants as young as 4 months use patterns of relative motion of the visible parts of a stimulus to determine whether this stimulus consisted of one or two objects. Similarly, Spelke et al. (1995)found that 4-month-olds analyzed the path of motion according to the principle that objects trace spatiotemporally continuous paths to determine the number of objects involved in an ambiguous event. In their studies, two screens are separated in space and objects emerged from the left side of the left screen and returned behind it, followed by an identical object emerging from the right side of the right screen and returned behind it. If no object appeared between the screens, infants inferred that two distinct objects must be involved in the event (Baillargeon and Graber, 1987; Xu and Carey, 1996).

Other types of physical knowledge also play a role in infants' object individuation. For example, Needham and Baillargeon (1997)showed that infants rely on support relations to resolve ambiguous displays. In one study, 8-month-old infants were shown a stationery display consisting of a box and a cylinder adjacent to each other. When the cylinder was suspended in mid-air (as opposed to lying on the floor of the apparatus) and appeared to be supported by the box, infants used the support relation to determine that they were parts of a single object: given that the cylinder was supported by one side of the box, it had to be connected to the box. Similarly, Needham and Baillargeon (1997)showed that knowledge of the solidity principle (one object cannot pass through another) can be used by 8-month-olds to disambiguate displays. If a thin piece of metal can be inserted between two adjacent parts of a display, infants use that information to determine that there are two objects.

Static configurational information also plays a role in how infants individuate objects, although these effects are generally weaker than those that rely on physical or spatiotemporal information and are modulated by the complexities of the objects and experience with particular objects (see Needham et al., 1997for a comprehensive review of these studies). In a series of studies, Needham (1998)found that when given a display of two featurally dissimilar objects adjacent to each other, a box and a hose, 7.5-month-old infants, but not 4.5- and 6.5-month-old infants, segmented the array into two objects. This result obtained when the hose was curved and the connecting surface between the two objects was only partially visible to the infants. Further, Needham showed that it was the complexity of the display which caused difficulties for the younger infants. When they simplified the display by using a straight hose and making the connection of the two objects completely visible to the infants, both 4.5- and 6.5-month-old infants parsed the array into two objects.

In another series of studies, Needham and Baillargeon (1998)demonstrated that infants can use yet another source of information in segmenting ambiguous displays into separate objects. Seeing part of an ambiguous display alone helps infants individuate adjacent objects. For example, after seeing the box standing alone on the stage for 5 s or after seeing the hose alone on the stage for 15 s, 4.5-month-old infants would parse the more complex array into two objects when shown the same box adjacent to the hose. Most strikingly, the effect of experience with the box helps the infants individuate objects even when the box was shown to the infants as young as 4.5 months in their homes 24 hours prior to the experiment.

It is clear from the literature reviewed above that young infants use many different strategies to solve the object individuation problem, and the resources they bring to bear are enhanced in various ways by learning during the first year of life. Whereas the studies by Needham and her colleagues certainly show sensitivity to perceptual differences among what might be parts of a single object or what might be two distinct objects (e.g. the blue, square part vs. the yellow, corrugated, curved part), none of the studies reviewed above addresses directly whether and when infants use object kind information in the context of object individuation.

By kind information, we mean information derived from classifying the stimuli according to antecedently represented categories in long-term memory. The distinction between kind representations (such as vehicle, animal, dog or chair) and property representations (such as blue, heavy, round or big) is not easy to draw – and it is an empirical question whether this distinction applies to infant object representations. If such a distinction can be sustained, kind representations may be seen to differ from property representations along some if not all of the following dimensions: kind representations are candidate meanings for count nouns which label object categories (Waxman and Markow, 1996); kind representations are stable, accessible, and long-term (Mandler, 1992); kind representations pick out functionally significant categories (Baldwin et al., 1993); and kind representations support category-based induction (Gelman and Markinan, 1986; Baldwin et al., 1993; Mandler and McDonough, 1996).

The objects used in the studies reviewed above were almost always fashioned into simple regular geometric shapes, and were not members of functional, namable kinds. Even Needham and Baillargeon's elegant box-hose studies did not actually use a box (it was a blue, rectangular object) or a hose (it was a yellow, corrugated, cylindrical object). There is considerable controversy concerning when infants first represent kind concepts. Visual habituation studies (e.g., Cohen and Younger, 1983; Quinn et al., 1993; Eimas and Quinn, 1994; Behl-Chadha, 1996) show that infants as young as 3 or 4 months habituate to discriminable stimuli from a single basic level or global basic category (e.g. dog or animal) and recover interest when shown stimuli of contrasting categories (e.g. cat or vehicle). However, these studies may reveal sensitivity to visual similarity alone, possibly abstracted during the habituation procedure itself. If infants have established long-term representations of kinds such as duck, car, cup, or ball, they should be able to use these representations to recognize objects as members of such kinds, and one would expect them to use these representations to help solve the object individuation problem.

Contrary to this expectation, Xu and Carey (1996)provide evidence that membership in kinds such as cup, bottle, ball, book, duck/animal, and truck/vehicle does not guide object individuation until the last few months of the first year. In these studies, ambiguity concerning the number of objects in the scene was introduced by showing the objects one at a time emerging from behind an occluder then returning behind the occluder. For example, in one of the experiments, 10-month-old infants were habituated to events in which a very familiar object, say a ball, appeared from behind a screen to one side of the stage and returned, then another very familiar object, say a bottle, appeared from behind the same screen to the other side of the stage and returned. The question was whether the infants would infer from the kind change that there must be two numerically distinct objects behind the screen. When the screen was removed, the infants were shown the expected outcome of two objects, namely a ball and a bottle, or the unexpected outcome of only one of the two objects, a ball or a bottle. Surprisingly, the 10-month-old infants did not look longer at the unexpected outcome: their pattern of looking did not differ from their baseline preference for two objects on baseline trials, as if there was not sufficient information to make a judgment. In contrast, when 10-month-old infants were given spatiotemporal information by simply showing them the two objects simultaneously for a brief 2 or 3 s at the beginning of the familiarization event, they looked longer at the unexpected outcome of one object when the screen was removed. It is not until 12 months of age that the majority of the infants in Xu and Carey's studies succeeded at using kind information alone to individuate objects.

Although various control experiments showed that Xu and Carey's method was sensitive, there are lingering doubts about the claim that infants do not use kind information in object individuation until about 12 months of age. To succeed in Xu and Carey's task, infants must be able to recall the representation of the first object, including details of its features, and compare this recalled representation with that of the currently visible second object in order to arrive at a representation of two distinct objects. None of the object segmentation studies reviewed above placed such episodic, short-term memory demands for recalling the details of a complex object and comparing them to those of a currently visible complex object.¹ Typically the stimuli were either completely and continuously visible or occluded only at the boundary between them. However, recent studies by Wilcox and Baillargeon (1998), in press) give some evidence that when a simpler paradigm is used to reduce information processing demand (event-monitoring as opposed to event-mapping, or event-mapping without the reversal of the trajectory), infants much younger than 10 months succeed in using the differences in object features to infer two distinct objects, under conditions of full occlusion of the first object when the second becomes visible. These objects, however, are simple shapes, not members of functionally relevant kinds.

The present study seeks to further test the hypothesis that infants do not use kind information for object individuation until about 12 months of age, using a paradigm that reduces the information processing load, in particular the short-term memory demands. Perhaps when the information processing demand is reduced in the experiment, infants would succeed earlier in using kind information for object individuation. We adapt the object segmentation paradigm of Spelke et al. (1993), using objects that clearly belong to namable kinds as opposed to nonsense shapes. Consider the following display: a toy duck with a flat bottom perching on the flat top of a toy car (Fig. 1). The duck has all the typical features of a duck, e.g. a duck's bill, eyes, distinct shape etc.; the car is a typical car with wheels, doors, metallic shine, etc. Adults perceive this display as unambiguous: we see a duck sitting on top of a car and we do not expect the two objects to be connected (this is confirmed in the preliminary experiment). Gestalt principles do not clearly specify the boundaries of the objects – cues such as violation of good continuation and good form are only weakly provided in such complex stimuli. While color changes between the duck and the car, it also changes between the car and the wheels, or the body of the car and its windows. Although the boundary between the duck and the car is marked by cues which specify distinct parts of an object, so too is the boundary between the top of the car and the rest of the car body, the car and its wheels, the head of the duck and the body of the duck, and so forth. Our knowledge about specific kinds of objects tells us that there are two objects in this array; perhaps because the perceptual evidence for distinct objects is ambiguous, we need to draw on our knowledge of object kind.

We use the duck-car display and ask whether 10-month-old infants can use object kind information to successfully individuate objects. The contrast between a toy duck and a toy car was chosen for several reasons. First, duck and car are basic level kinds, in Rosch and Mervis's sense (Rosch et al., 1976) and many have assumed that the infants' first kinds are basic level kinds (Macnamara, 1987). Second, others have argued that infants begin with more global kind distinctions, such as bird vs. vehicle, or animal vs. vehicle (Mandler, 1992; Mervis, 1987), and duck and car also contrast as these levels. Finally, under the assumption that infants' first words for objects map onto kind concepts, the fact that `duck' and `car' are among children's early words (Fenson et al., 1991) suggests that they may be among the first kind concepts infants represent. Thus we are probing whether infants already represent kinds such as duck and car prior to learning words for them.

In Experiment 1, the infants are habituated to a display that consists of a duck perching on top of a car with a hand suspended right above the display without touching it. During the test trials, the hand grasps the head of the duck and lifts it up. In the expected outcome, the duck is lifted but the car stays in place. In the unexpected outcome, the car is lifted along with the duck as if they were parts of one complex object. If infants have segmented the display into two objects, they should look longer at the unexpected outcome in which the car is lifted along with the duck. It is important to note that both objects are in plain view throughout the experiment, so the infants will not have to actively retrieve their representation of the first object upon seeing the second in order to arrive at a representation of two distinct objects. If the findings of Xu and Carey (1996)reflect the inability of 10-month-old infants to use kind information for object individuation, they should fail to use the differences between the toy duck and the toy car to set up a representation of two distinct objects, whereas 12-month-old infants will succeed. Given Needham and Baillargeon's findings that experience with one part of an ambiguous stimulus helps in the object individuation problem, we include a condition in which infants are shown the duck alone, and allowed to handle it, as well as the car alone, and allowed to handle it, prior to habituation to the duck-on-the-car. If infants truly lack antecedent duck/animal or car/vehicle kind representations, then such a small amount of experience might not be sufficient to help them in the object individuation task. In contrast, if infants do represent these kinds antecedently, the representation should be primed and would be expected to facilitate individuating the duck from the car in the ambiguous duck-car display.

Before beginning the infant studies, a preliminary study tested adults' perception of the display as two distinct objects.