Infants’ use of contextual cues in the generalization of effective actions from imitation

Journal of Experimental Child Psychology 116 (2013) 510–531

Contents lists available at SciVerse ScienceDirect

Journal of Experimental Child Psychology journal homepage: www.elsevier.com/locate/jecp

Infants’ use of contextual cues in the generalization of effective actions from imitation Dahe Jessica Yang a, Emily W. Bushnell a, David W. Buchanan b, David M. Sobel c,⇑ a

Department of Psychology, Tufts University, Medford, MA 02155, USA IBM Research Division, IBM T. J. Watson Research Center, P.O. Box 704, Yorktown Heights, NY 10598, USA c Department of Cognitive, Linguistic, and Psychological Sciences (CLPS), Brown University, Providence, RI 02912, USA b

a r t i c l e

i n f o

Article history: Available online 26 January 2013 Keywords: Imitation Rational categorization Causal reasoning Infancy

a b s t r a c t We examined infants’ ability to generalize effective actions in an imitation task. In Experiment 1, 15-month-olds imitated effective and ineffective actions on two similarly designed toys. They were then shown a third toy of the same design with both actions available. Children reliably touched and manipulated the effective action handle first and more persistently. In Experiment 2, however, 15-month-olds did not generalize the efficacy of the action when the test toy was different from the two demonstration toys. Experiment 3 replicated the findings of Experiment 2 but also showed that infants generalized efficacy when the demonstration toys differed from one another as well as from the test toy. Our findings are consistent with a computational model that uses certain rational pedagogical assumptions. Overall, the results suggest that 15-month-olds are sensitive to the sampling information they observe and use this information to guide whether to generalize efficacy information they learn from imitation. Ó 2012 Elsevier Inc. All rights reserved.

Introduction Infants reproduce the behaviors of others from very early ages (e.g., Meltzoff & Moore, 1977, 1983), and by the end of the first year they readily imitate simple goal-directed actions with novel objects (Elsner & Aschersleben, 2003; Herbert, Gross, & Hayne, 2006; Meltzoff, 1988a). During the second

⇑ Corresponding author. Fax: +1 401 863 2255. E-mail address: [email protected] (D.M. Sobel). 0022-0965/$ - see front matter Ó 2012 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.jecp.2012.09.013

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

511

year, they also imitate multistep activities (Bauer & Hertsgaard, 1993; Herbert & Hayne, 2000), toolusing strategies (Bushnell & Boudreau, 1998; Chen & Siegler, 2000), and even inefficient (Lyons, Young, & Keil, 2007; Nagell, Olguin, & Tomasello, 1993) or highly unusual actions (Meltzoff, 1988b). As Tomasello (1999) and others have established, observational learning is clearly a prevalent and important mechanism by which young children efficiently acquire cultural skills and learn the conventional uses of objects. Imitation also allows children the opportunity to learn the efficacy of objects and actions toward them—that is, what actions on objects (and the objects themselves) are capable of producing in the way of effects. This is particularly true in cases where the action required on an object produces an effect incommensurate with the nature of the actions required to bring about that effect. For instance, one might look at a hammer and recognize its affordances, but other artifacts lack such transparent efficacy. We can press a button to activate a teakettle, but there is nothing about pressing a button that should cause water to boil. Moreover, once we learn that pressing the button causes the water to boil, we can generalize that knowledge to the efficacy of novel kettles (which might be activated in a variety of ways) and, more interesting, to novel buttons (which might have a variety of effects). Although there are sophisticated descriptions of learning mechanisms that consider how children learn this kind of information from their own observations and interventions on the environment (e.g., Gopnik et al., 2004), most of this knowledge is socially constructed (e.g., Boyd, Richerson, & Henrich, 2011; Harris & Koenig, 2006). Therefore, it is critical to know the extent to which infants can learn action–effect relations from others and how generally they will apply that knowledge. Our goal in this article is to examine how children engage in the process of generalizing efficacy learned from imitation. Generalizing an action’s efficacy from imitation to additional objects is a crucial capacity. Without it, an infant’s knowledge of how to ‘‘make things work’’ is limited to the collection of objects the infant happens to witness being acted on; with it, imitation becomes generative and the infant’s action knowledge may grow exponentially beyond the initial instances observed (see, e.g., Bushnell, Sidman, & Brugger, 2006). Thus, we ask the following questions. When infants imitate another’s actions on an object that have been shown to produce a novel effect, what do they know about that object’s efficacy? Are they just reproducing the action, or do they also note its consequences and expect their own action to likewise generate an effect? If the latter is true, do infants understand the action to produce an effect regardless of the object on which it is performed, or do they link the action’s efficacy only to the specific object involved in the original imitation? We first review literature on infants’ imitative abilities as indicating an understanding of the efficacy of action they observe and an expectation of their own action to be effective. Second, we consider the specificity of infants’ generalization capacities by couching the problem of generalization as one of category membership. We then present three experiments, manipulating how likely it is that objects are members of the same category (and thus might share efficacy). Following the experiments, we present a computational model that treats categorization as a rational process. We show the model’s fit with the current data and possible extensions of this model. Our goal in presenting this model is to illustrate how the way infants treat the category membership of the objects affects whether they expect what they learn from others’ efficacious action to generalize to other objects. Finally, we discuss these data in terms of our categorization-based approach as well as alternative accounts, particularly the case in which the generalization process is part of the representation of the action children perform and not a categorical inference they make. It is important to clarify here that when we speak of an infant’s learning or understanding an action’s efficacy, we mean only that that the infant expects the action to lead systematically to an effect of some sort. In most such cases, infants clearly do not understand the mechanical, electrical, or other processes by which the action literally causes the effect to ensue. Thus, we distinguish between understanding efficacy and understanding causality; accordingly, we use the term efficacy throughout the rest of this article. What do infants learn from imitation? There are numerous investigations suggesting that infants are not just reproducing action when they imitate (i.e., ‘‘mimicking’’ or imitating ‘‘blind’’; cf. Want & Harris, 2002). For instance,

512

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

18-month-olds performed an action they inferred a model meant to do rather than the failed attempt the model had actually demonstrated (Meltzoff, 1995), and 14- to 18-month-olds imitated actions the model verbally marked as intentional more frequently than actions she marked as accidental (Carpenter, Akhtar, & Tomasello, 1998). Similarly, Gergely, Bekkering, and Király (2002) showed that infants imitated an unusual strategy for turning on a light only when the model apparently chose to do it that way; when infants could see that the model was constrained from using a more conventional strategy, they accomplished the task in the more usual manner rather than as the model had done it. There are also numerous investigations suggesting that infants note the effectiveness or efficacy of the action demonstrated to them. For instance, Bauer (1992) found that infants shown multistep sequences imitated ‘‘enabling’’ actions vital to the goal more frequently than they imitated irrelevant actions embedded in the sequence. Similarly, Brugger, Lariviere, Mumme, and Bushnell (2007) found that infants imitated a particular action more frequently when it was necessary (by virtue of its location) to achieve a goal than when it was not. In addition, 18-month-olds register the consistency of the relation between an action and an outcome; they imitated an action more precisely when it always produced an effect compared with when it only sometimes produced an effect (Schulz, Hooppell, & Jenkins, 2008). Furthermore, infants have some understanding that their own action will also generate the effect they observed another action to generate. For instance, 12-month-olds, but not younger infants, looked to see whether their imitated action would produce an effect as the model’s action had (Carpenter, Nagell, & Tomasello, 1998). These findings, however, were based on infants understanding a single action. Elsner and Aschersleben (2003) investigated whether infants connected particular actions and outcomes within the context of imitation. Infants were shown demonstrations of two actions (pressing and pulling) on the same object, each of which led to a distinct effect. Infants were then given a turn with the object with either the same action–effect contingencies or those contingencies reversed. Older infants (15- and 18-month-olds) responded differently in the two contingency conditions; they acted on the objects more frequently when the contingencies were the same as in the demonstrations. Younger infants (12-month-olds) did not distinguish between the contingency conditions, although they did perform the target actions more frequently than infants not shown the demonstrations at all. The younger infants appeared to be mimicking the model’s actions, whereas the older infants connected each observed action with a specific effect. Although there is some development based on the physical distance between the action and the effect (Bushnell et al., 2006; Meltzoff, Waisman, & Gopnik, 2012), it would appear that by 15 months infants register something about the relation between an action and its effect. In Elsner and Aschersleben’s (2003) research, however, the two actions and their effects were demonstrated on one and the same object. Thus, it is not clear whether infants linked the actions and effects together only in the context of that object or whether they generalized the efficacy of the modeler’s actions and would apply them to other objects. Elsner and Pauen (2007) examined whether 12- and 15-month-olds would generalize efficacy to other objects. They showed infants three objects: two objects with different part configurations and an effect toy. The configurations differed such that one object could be inserted into the effect toy and the other object could not. When the object that could be inserted into the effect toy touched it, the effect toy produced a novel sound. When the other object contacted the effect toy, it did not make a sound. The 12-month-olds rarely imitated these actions, whereas the 15-month-olds did so more frequently and used the effective object more often than the ineffective object. At the end of their procedure, the authors showed children six novel objects: three with the efficacious part configuration and three with the inefficacious part configuration. The 12-month-olds played with these toys equally (and not often), whereas the 15-month-olds played with the objects with the efficacious parts more often. Thus, these data suggest that infants’ generalization abilities are developing between 12 and 15 months. Critically, Trauble and Pauen (2007) showed that 12-month-olds categorized these stimuli differently when infants observed their efficacy. They showed infants the objects used during the generalization phase of Elsner and Pauen’s (2007) study. When the functions of the objects were demonstrated (i.e., objects with one part configuration activated a similar, but not identical, ‘‘effect box’’, whereas objects with the other configuration did not), infants treated the objects with the same configuration of parts as members of the same category (measured by lower examination time spent

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

513

with each object, following Oakes, Madole, & Cohen, 1991). Given that 12-month-olds register objects with similar efficacious parts as members of the same category, what appears to be developing between 12 and 15 months of age is infants’ generalization capacities. Generalization as a rational process: overview of the experiments There are two open questions from the experiments by Elsner, Pauen, and colleagues. The first is that because the actions infants observed were the same, whether an action was effective given a particular object was determined by the configuration of that object’s critical part (i.e., whether it had the open end of the T-joint, so that it could be inserted into the effect box, or had the closed end, so that it would not fit). It is possible that infants learned nothing about producing effects per se but instead learned an ‘‘affordance’’ for each object’s manipulandum that was similar between the demonstration and test objects. That is, they learned that objects with a particular configuration could be placed in the goal object, whereas objects with the other configuration could not, much in the same way as infants could simply learn by observation that buttons are supposed to be pressed. There is some evidence for this interpretation. Yang, Sidman, and Bushnell (2010) found that 15-month-olds transferred a target action (i.e., pressing a button) to a novel object even after they had imitated it only on ‘‘naked’’ handles—that is, on buttons that were not connected to any object and whose pressing had led to no audible or visible effect at all (Study 4). Thus, it is possible that infants do not learn about the relation between actions and effects from imitation but rather just learn to act based on an object’s affordances. The second remaining question concerns the nature of the generalization ability exhibited in earlier experiments: is it rational? Many have suggested that infant imitation is rational (e.g., Csibra & Gergely, 2009), as indicated by their capacities to integrate the intentionality of the physical causal structure of the situation into account when acting (based on much of the evidence presented above). We hypothesize that infants’ generalization ability is likewise rational insofar as infants’ generalization is a function of their understanding of category membership. To test this, we adapted the rational model of categorization (RMC) (Anderson, 1990; Sanborn, Griffiths, & Navarro, 2010) to construct a computational model that explains the current data. We chose this framework because it allows us to specify what information in the environment would be critical for the inference as well as to optimally choose what course of action to take given different datasets. Our goal was to see whether this computational framework would make similar predictions about generalization based on category information as infants in our experiments as well as whether it would make novel predictions that we could test with our dataset. We specifically hypothesized that what 15-month-olds understand (that potentially 12-month-olds do not, based on the work by Elsner & Pauen, 2007) is that if objects are in the same category, they should possess the same efficacy. In these experiments, 15-month-olds were shown that a particular action on a toy led to an effect while a different action on another toy produced no effect. Infants then interacted with a third toy that afforded both actions, and we observed whether they preferentially performed the effective action. In Experiment 1, all three of these toys were the same in appearance except for the manipulables (handles). In Experiments 2 and 3, the identity of the test toy or of both the demonstration and test toys varied across the trials. This allowed us to determine whether and under what conditions infants generalized any knowledge of efficacy they learned from imitating the demonstrated actions. We hypothesized that construing all three objects as members of the same category would result in learners being more likely to generalize the efficacy observed in the demonstration to the test object. On the other hand, if learners categorize the test object as different from the demonstration objects, they would be less likely to generalize and hence would act randomly on the test object. Thus, we predicted that when the three objects are identical (Experiment 1), generalization should be quite strong. In contrast, when the two demonstration objects are the same but are different from the test object, generalization should be less likely (Experiments 2 and 3). Finally, we predicted that when all three objects are different in appearance (Experiment 3), generalization should be moderately likely because now they might be construed as members of a single broad category. This hypothesis is consistent with research indicating the role of multiple exemplars in promoting infants’ generalization in other cognitive domains (e.g., Cohen & Strauss, 1979; Liu, Golinkoff, & Sak, 2001; Quinn & Bhatt, 2005;

514

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

Younger & Fearing, 2000). For instance, Younger and Fearing (2000) showed that 10-month-olds were capable of registering multiple basic-level categories based on the generality of the stimuli they observed; in contrast, younger infants seemed to register that all stimuli were members of a single category regardless of their diversity. These predictions come from making a particular assumption about imitation—that children interpret it as a social and didactic act. That is, children infer that the demonstration objects are being shown to them for a reason based on the experimenter’s intentions. Preschoolers’ understanding of the intentions of a modeler clearly influences the inferences these children make from that individual’s actions and the subsequent results (e.g., Bonawitz et al., 2011; Sobel & Sommerville, 2009). Gweon, Tenenbaum, and Schulz (2010) showed that 15-month-olds also appear to be capable of making causal inferences based on what is demonstrated. The studies cited above suggest that children make inferences about a modeler’s intentions. This point is critical because when we build our computational model, we can contrast the predictions of a model that makes this assumption with other alternatives for generalization (which we consider in the Discussion section). Experiment 1 In the first experiment, 15-month-olds were introduced to two objects that looked alike but had different handles, each affording a different action (i.e., pressing down or pulling out). Manipulating one of the handles produced an interesting audio–visual effect, whereas manipulating the other handle produced no such effect. After infants observed and imitated both the effective and ineffective actions, they were shown a third ‘‘hybrid’’ object of the same design but with both handles available. We observed which handle children first touched and acted on. If children register that an action produces an effect, they should act on the handle they previously saw as effective. If they are just mimicking movement patterns or learning affordances without appreciating each action’s efficacy, they should be equally likely to act on either handle. Method Participants The final sample consisted of 16 healthy full-term 15-month-olds (9 girls and 7 boys, mean age = 14.87 months, range = 14.00–15.97). An additional 11 infants were tested but not included in the final sample because of experimenter error (n = 2), because they were fussy or uncooperative in general (n = 4), or because they failed to produce at least one of the demonstrated actions during the learning trials (n = 5). Infants were recruited from nearby communities by mail solicitation to new parents. Their ethnic distribution reflected the demographics of these communities; as such, the sample was predominantly Caucasian, although other ethnicities and races were also represented. Setting and stimulus materials The parent, infant, and experimenter sat around a 45  90-cm wooden table. The experimenter and parent sat opposite each other at the short ends of the table, and the infant sat in a booster seat fastened to one of the long sides of the table. A video camera was placed between the baby and the experimenter. An assistant operated the camera and brought the stimulus toys out to the experimenter from behind a small partition. Two sets of novel toys were constructed for the procedure (referred to here as the bunny toys and cow toys [see Figs. 1A and B], although these labels were never presented to participants). Each set consisted of three toys similar in shape, size, and color but with different handles protruding from the toys’ bodies. Within each set, one toy had a button that could be pressed down (the press toy), another had a knob that could be pulled out (the pull toy), and another toy had both of these handles protruding—one to the left and the other to the right (the hybrid toy). Each of the three bunny toys (Fig. 1A) was a pink plastic cylinder (16 cm in diameter  16 cm in height) with a hinged lid. The lid had a circular transparent window so that one could see into the cylinder, which contained a white stuffed bunny. When the toy was activated, the lid opened and the

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

515

Fig. 1. Bunny toy (A) and cow toy (B) used throughout the experiments. The single-activator objects were used in all of the experiments. The dual-activator object was used in Experiments 1 and 2.

bunny rose partway out of the cylinder and made a 3-s ‘‘yahoo’’ sound effect. The press version of the bunny toy had a curved base (8  6  3.5 cm) protruding at the bottom with a red cylindrical pressbutton (3 cm in diameter  4 cm in height) on the base. The pull version had a metal rod (4.5 cm in length) protruding from the lower portion with a yellow spherical pull-knob (3 cm in diameter) at its end. When fully pressed, the button traversed approximately 1.5 cm downward, and when fully pulled, the knob traversed approximately 2.2 cm farther out; both the button and the knob returned to their original positions after they were released. Neither the button nor the knob moved apart from the object in any other direction if a force other than pressing or pulling was applied to it. The third hybrid bunny toy had both the press-button and the pull-knob protruding from its body. These were set apart from each other by 90°, and their positions could be interchanged so that their left/right locations could be varied across infants. The press and pull toys had hidden on/off switches, and in the procedure for any given infant, one toy was turned on so that moving its handle produced the effect while the other toy was turned off so that moving its handle did not yield the effect. The hybrid toy did not produce an effect regardless of which handle was manipulated. The cow toys (see Fig. 1B) were designed in a similar manner. Each was a square or rectangular box covered with light blue foam. The press version of the cow toy measured 17  14  18.5 cm and had an L-shaped rod (4  4 cm) protruding from its lower front with the red cylindrical press-button on

516

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

the end. The pull version of the cow toy measured 16.5  20.5  12 cm and had a metal rod (4.5 cm in length) protruding from its lower front with the yellow spherical pull-knob at the end. When the button was pressed or the knob was pulled, a hidden sounding-can inverted and the toy produced a loud drawn-out mooing sound (3 s in length). The hybrid cow toy was a 16.5-cm cube and had both the press-button and the pull-knob set apart from one another at a 90° angle; as with the hybrid bunny toy, the positions of these could be interchanged. Like the bunny toys, the press and pull cow toys could be set to produce the effect or not when manipulated, and the hybrid toy did not yield the effect regardless of which handle was manipulated. Procedure On arrival, the infant and the experimenter played in a waiting area while the parent filled out consent materials nearby. The parent and infant were then led into the testing room and seated at the table as described above. The parent was asked to remain quiet and nondirective throughout the experiment but was permitted to provide emotional support and encouragement (e.g., by nodding and smiling) if the infant solicited the parent’s attention. Two warm-up trials were initially conducted to establish rapport with the infant and to familiarize him or her with the routine of receiving toys from the experimenter one after another. On each warmup trial, an assistant brought a toy out from behind the partition and placed it in front of the experimenter, who remarked on it enthusiastically and then gave it to the infant. The infant was allowed to play freely with the toy for approximately 60 s, and then the assistant brought out the next warm-up toy and simultaneously took the first toy away. Infants’ attention was typically attracted to the new object, and so they relinquished the prior object readily. The warm-up toys were a clear cylinder with beads inside and a pair of colorful blocks. These afforded behaviors such as shaking, banging, and rotating, but they did not have manipulable parts akin to the buttons or knobs on the stimulus toys and they did not yield any special effects when acted on in a particular manner. After the infant’s turn with the second warm-up toy, the formal experimental procedure began. Half of the infants were shown the set of bunny toys one by one, and half were shown the set of cow toys. The infant was first shown one of the single-handled toys in the appropriate set; half of the infants were shown the press toy first, and half were shown the pull toy first. This toy was initially held out of the infant’s reach while the experimenter engaged the infant by saying, ‘‘Wow, look! Look at what I can do!’’ and then performed the appropriate action on the toy (pressing or pulling its handle). The first toy either was effective, and thus produced its effect, or was ineffective, and thus did not yield an effect; half of the infants with each toy set saw an effective toy first, and half saw an ineffective toy first. The experimenter performed the action in question three distinct times and then gave the toy to the infant, who was allowed to play with the toy until the target action had been executed several times and for at least 30 s. Next, the other single-handled toy from the same set was brought out and the first toy was set aside out of the infant’s reach (but still visible). If the first toy had produced the bunny or cow effect, the second toy did not and vice versa. As on the first trial, the experimenter solicited the infant’s attention, performed the appropriate action on the toy three distinct times, and then gave it to the infant for approximately 30 s of free play. The entire sequence of demonstrations by the experimenter and free play turns for the baby with each of the two single-handled toys was then repeated, so that the infant observed and played with both the effective toy and the ineffective toy twice each. If the infant failed to produce each of the target actions (i.e., pressing and pulling) during at least one of the free play turns with the corresponding toy, the infant was not included in the final analyses. This criterion ensured that all infants included in the analyses were familiar with and had the motor capacity to produce both of the actions. Of the excluded infants, 5 were excluded on the basis of this criterion; surprisingly, all 5 infants failed to act on the effective toy. A couple of these infants seemed wary of the cow mooing sound, and the others persisted in attempting to move the handle in impossible ways (e.g., lifting the press button, twisting the pull knob). After the sequence of demonstrations and baby’s turns with the single-handled toys, the third hybrid toy with both handles on it was brought out and the other two toys were removed from the table. The hybrid toy was placed in front of the infant with the handles both in reach and symmetrically positioned, so that one handle was on the infant’s left and the other was on the right; the left/right

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

517

position of the handles was determined randomly across infants. The experimenter did not perform any actions on this test toy. Instead, the experimenter simply offered the toy to the infant and called attention to it, saying, ‘‘Heh, how about this one?’’ The infant was then allowed to play with the hybrid toy for approximately 30 s. After the test trial with the hybrid toy was over, the parent and infant were given a small gift of appreciation and escorted from the laboratory. Coding Infants’ activities with the toys on the several experimental trials were coded from video records on a second-by-second time line representing the 30 s infants had to play with each toy. A pair of student research assistants, working together, viewed the videotapes of all infants. For the demonstration trials with single-handled toys, the scorers coded how many seconds infants spent touching each toy’s handle and also how many seconds they spent actively manipulating the handle. Manipulating was defined as pressing the button down or pulling the knob out to nearly its full excursion—that is, moving the handle sufficiently to have produced the toy’s effect if the toy were active. Touching encompassed manipulating but was a larger category that also included just resting the hand or fingers on the handle and moving the handle in ways that did not (could not) produce the effect (e.g., twisting the button, wagging the pull-knob from side to side). For the test trial with the hybrid toy, the scorers coded the durations of touching and manipulating for each one of the two distinct handles, and they also coded which of the two handles the infant touched and manipulated first. A second team of two different observers rescored a random sample of 25% of the data tapes independently to assess reliability for scoring the target behaviors. Overall agreement regarding the durations for touching and manipulating on the demonstration and test trials was 96% (kappa = .92). Overall agreement regarding which handle was first touched and first manipulated on the test trial was 100%. The observers could not be ‘‘blind’’ regarding which handle was functional on the test trial because they also scored the demonstration trials. However, the observers were naive regarding the overall purpose of the experiment, and the target behaviors were overt, well defined, and scored with high interrater reliability. Results Recall that to be included in the final analyses, infants needed to execute both pressing the button and pulling the handle at some point during the demonstration trials. We first present analyses of the extent to which infants engaged in these actions during the demonstration phase (i.e., the extent to which they imitated) and then present analyses of their generalization during the test phase. All analyses for both the demonstration and test phases employed one-tailed tests of significance because one would expect infants to act preferentially toward the working handle; a preference for the ineffective handle would be inexplicable. Demonstration phase Table 1 shows the durations of time infants executed the target actions on the effective handle (i.e., the handle that produced the effect) versus the ineffective handle (i.e., the handle that did not). Although on average infants spent more time manipulating the effective handle (M = 12.56 s) compared with the ineffective handle (M = 8.25 s), this difference was not statistically significant, t(15) = 1.44, p = .09. Considered individually, 8 infants spent more time manipulating the effective handle, 6 infants spent more time manipulating the ineffective handle, and 2 infants spent the same amount of time manipulating the two handles. This distribution of individuals was not different from that expected by chance, binomial test, p = .40. Thus, from the demonstration trials, infants were familiar with both handles and their target actions, and they had roughly equivalent practice with the effective and ineffective actions. Test trial Table 2 shows the numbers of infants who first touched and first manipulated each of the handles on the test toy. Most infants first touched and also first manipulated the handle that had been effective during the demonstration trials. Of the 14 infants who initially touched a single handle, 11 touched the

518

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531 Table 1 Mean durations (in seconds) of manipulating the different kinds of handles during the demonstration phase of each experiment. Effective

Ineffective

Experiment 1 Mean (SD)

12.56 10.18

8.25 5.48

t = 1.44

Experiment 2 Mean (SD)

7.93 6.42

9.5 7.55

t = –0.58

Experiment 3: Same condition Mean 12.96 (SD) 13.1

14.58 10.88

t = –0.56

Experiment 3: Varied condition Mean 11.87 (SD) 10.5

15.46 12.51

t = –1.05

Table 2 Numbers of infants who first touched and first manipulated the effective and ineffective handles on the test toy across the experiments. Effective

Ineffective

Other response

Experiment 1 (N = 16) Touched first Manipulated first

11 12

3 3

2 1

Experiment 2 (N = 16) Touched first Manipulated first

8 9

8 5

0 2

Experiment 3: Same condition (N = 24) Touched first Manipulated first

10 10

10 12

4 2

5 7

3 0

Experiment 3: Varied condition (N = 24) Touched first 16 Manipulated first 17

Note: ‘‘Other response’’ refers to an infant’s touching both handles simultaneously or manipulating neither handle on the test trial. Table 3 Mean durations (in seconds) infants spent touching and manipulating the effective and ineffective handles on the test trial of each experiment. Effective handle

Experiment 1 Touching Manipulating

(SD)

Mean

(SD)

16.00 7.25

1.64 1.44

9.38 2.25

1.48 0.6

t = 3.46*** t = 2.84*

2.21 1.38

12.88 5.5

2.14 1.57

t = 0.10 t = 0.42

1.5 1.37

13.83 5.63

1.91 1.37

t = 0.51 t = 0.14

1.43 1.09

11.17 4.67

1.61 1.1

t = 0.80 t = 1.12

Experiment 2 Touching 12.63 Manipulating 4.56 Experiment 3: Same condition Touching 14.83 Manipulating 5.88 Experiment 3: Varied condition Touching 13.08 Manipulating 6.96 *

p < .05 p < .01

***

Ineffective handle

Mean

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

519

effective handle first; this distribution of individuals is significantly different from that expected by chance, binomial test, p = .03. (The 2 infants who initially touched both handles at the same time were considered as ‘‘ties’’ and thus excluded for the binomial statistic.) Similarly, 12 of the 15 infants who manipulated at least one of the handles on the test trial manipulated the effective handle first. This distribution is also significantly different from that expected by chance, binomial test, p = .02. All 3 infants who initially manipulated the ineffective handle chose to press on the hybrid test toy. The mean durations of time infants spent touching and also manipulating each of the two handles on the hybrid toy during the test trial are shown in Table 3. Recall that neither action actually activated the test toy. Nevertheless, infants spent more time touching the handle that had been effective during the demonstrations (M = 16.00 s) than touching the handle that had not yielded an effect before (M = 9.38 s), t(15) = 3.46, p = .002, Cohen’s d = 4.23. Similarly, they manipulated the handle that had been effective during the demonstrations for longer durations (M = 7.25 s) than they manipulated the handle that had not yielded an effect before (M = 2.25 s), t(15) = 2.84, p = .006, Cohen’s d = 4.50. All infants touched each of the handles for at least some time during the test trial, but 5 infants manipulated only the previously effective handle, 1 infant manipulated only the previously ineffective handle, and 1 infant did not manipulate either handle. Discussion Experiment 1 showed that 15-month-olds generalized the efficacy of actions on objects they imitated to a similar novel object. When given a choice between a handle whose movement had produced an effect earlier and a handle whose movement had not, infants systematically reached for and operated the effective handle first. This preference could only have been based on infants perceiving and remembering the consequences of moving each handle because all other variables that might determine preference (e.g., left/right position, pressing vs. pulling, first vs. second action practiced) were counterbalanced across the working/broken distinction. Thus, when 15-month-olds imitate a modeler’s goal-directed action, they are not simply mimicking the action or exploring the object’s ‘‘action affordance’’ (i.e., how to move it). Moreover, infants acted more on the effective handle than on the ineffective handle throughout the test trial even though neither handle led to an effect during that trial. Schulz and colleagues (2008) argued that the amount of time infants persist with an action might index their expectations about the efficacy of the action. Infants’ persistence with the effective handle further documents that they perceived and remembered what happened when it was manipulated during the demonstration trials. This conclusion is consistent with Carpenter and Akhtar et al. (1998) observation that 12-month-olds ‘‘looked for’’ an effect when they reproduced a model’s goal-directed action. It is also consistent with Elsner and Aschersleben’s (2003) finding that 15- and 18-month-olds were sensitive to specific pairings of actions and effects. The question remains, however, whether infants understand the efficacy of the one action as specific to this class of objects or as general to all objects with this kind of handle. We suggest that the perceptual similarity across the familiarization and test objects indicates that the three objects were all members of the same category and, thus, generalization of the effective action is likely, as indeed we observed. In Experiment 2, infants were tested with a procedure similar to that of Experiment 1 except that the test object was different in kind from the objects during the demonstration phase. If infants are learning that the efficacy of actions is independent of the object, they should generalize in this condition. Alternatively, if infants construe the test object to be from a different category than the familiarization objects, they should be unlikely to generalize. Because of the dissimilarity between the demonstration and test objects in Experiment 2, we predicted the latter result. Experiment 2 In the second experiment, 15-month-olds were presented with the same demonstrations and opportunities to imitate as in Experiment 1 and then were offered a hybrid test object with both of the demonstration handles. In this second experiment, the test object differed in overall appearance

520

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

from the two demonstration toys. If infants treat the demonstration as indicating that the action on the effective handle is generally effective, they should choose that handle on this novel test object as in Experiment 1. However, because the test object is distinct, infants may categorize it differently from the demonstration toys and thus, according to the RMC, would be less likely to generalize the effective action to the novel toy. Method Participants The final sample consisted of 16 healthy full-term 15-month-olds (9 girls and 7 boys, mean age = 15.03 months, range = 14.17–15.97). An additional 9 infants were tested but not included in the final sample because of experimenter error (n = 2), because they were fussy or uncooperative in general (n = 3), or because they failed to produce at least one of the demonstrated actions during the learning trials (n = 4). Infants were recruited in the same manner and from the same communities as in Experiment 1, but none of them had participated in Experiment 1. Setting and stimulus materials The setting and stimulus materials used in Experiment 2 were the same as those used in Experiment 1. Procedure The procedure for Experiment 2 was identical to that for Experiment 1 except that the hybrid test toy came from the alternative set of toys to those used in the demonstrations instead of from the same set. Thus, if infants saw the single-handled demonstration toys from the bunny set, the test toy came from the cow set and vice versa. In this way, we investigated whether infants generalized the expectation of an outcome they learned from acting on a particular handle in the demonstrations to a novel test object from a putatively different category. Coding The same coding system as described for Experiment 1 was used in Experiment 2. All trials were scored by two observers working together, and reliability was again examined by having another pair of observers independently score a random sample of 25% of the data. Overall agreement regarding the durations for touching and manipulating on the demonstration and test trials was 97% (kappa = .93). Overall agreement regarding which handle was first touched and first manipulated on the test trial was 100%. Results As in Experiment 1, to be included in the final analyses, infants needed to execute both pressing the button and pulling the handle at some point during the demonstration trials. We again first consider imitation during the demonstration phase and then generalization during the test trial. Demonstration phase The durations of time infants executed the target actions on the effective and ineffective handles in Experiment 2 are displayed in Table 1. On average, infants spent 7.93 s manipulating the effective handle and 9.5 s manipulating the ineffective handle; the difference between these durations was not statistically significant, t(15) = 0.58, p = .29. Considered individually, 7 infants spent more time manipulating the effective handle and 9 infants spent more time manipulating the ineffective handle. This distribution of infants is not different from that expected by chance, binomial test, p = .40. Thus, as in Experiment 1, from the demonstration trials, all infants were familiar with both handles and their actions, and they had practiced the effective and ineffective actions equivalently.

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

521

Test trial The numbers of infants who first touched and first manipulated each of the handles on the test toy are shown in Table 2. Here, 8 of the infants touched the handle that had been effective during the demonstrations first, and 8 touched the handle that had been ineffective first; this distribution is obviously not different from that expected by chance. Similarly, 9 infants manipulated the effective handle first and 5 manipulated the ineffective handle first; this distribution is also not different from that expected by chance, binomial test, p = .21. The mean durations of time infants spent touching and also manipulating each of the two handles during the test trial are shown in Table 3. As can be seen, infants in Experiment 2 did not act more persistently on one handle than on the other. They touched the handle that had previously yielded an effect for an average of 12.63 s and touched the handle that had not yielded an effect for an average of 12.88 s, t(15) = 0.10, p = .46. Similarly, they manipulated the handle that had previously yielded an effect for an average of 4.56 s and manipulated the handle that had not yielded an effect for an average of 5.50 s, t(15) = 0.42, p = .34. All infants touched each of the handles for at least some time during the test trial, but 4 infants manipulated only the previously effective handle, 3 manipulated only the previously ineffective handle, and 2 did not manipulate either handle. Thus, in Experiment 2 infants did not favor the effective handle on the test trial in any respect, whereas in Experiment 1 infants had touched the effective handle first and for longer durations and had manipulated it first and for longer durations. The two experiments were identical except that in Experiment 2 the test toy was different in kind from the demonstration toys rather than of the same kind; hence, we compared the results of the two studies directly. The distributions of individual infants who first touched the effective handle versus the ineffective handle on the test toy (see Table 2) were not significantly different between the two experiments (Fisher’s exact test, p = .14), nor were the distributions of infants who first manipulated the effective handle versus the ineffective handle (Fisher’s exact test, p = .43). A 2 (Experiment: 1 vs. 2)  2 (Handle: effective vs. ineffective) mixed analysis of variance (ANOVA) on the durations of touching the test toy’s handles (see Table 3) did show a significant interaction between experiment and handle, F(1, 30) = 4.55, p = .041, g2 = .05. A similar ANOVA on the durations of manipulating the test toy’s handles showed the same significant interaction, F(1, 30) = 4.35, p = .046, g2 = .08. There were no main effects in either of these analyses; they yielded only the significant Experiment  Handle interactions, which are well explained by infants’ touching and manipulating the effective handle for longer than the ineffective handle in Experiment 1 but interacting with the two handles for equal durations in Experiment 2. Discussion Experiment 2 revealed no evidence that infants generalize their knowledge of an action’s efficacy to an object distinct from those with which they learned the action. The infants presumably learned that one action was effective and the other was not with the demonstration toys because their demonstration experience was identical to that of infants in Experiment 1. However, when the test toy differed in appearance from the demonstration toys, 15-month-olds did not show a preference for the handle and action that had been effective with the demonstration toys as infants in Experiment 1 had shown. Instead, they chose randomly between the test toy’s two handles and touched and manipulated them equally. They behaved as if they had no hypotheses about what actions might be effective or ineffective in the context of a new object. One could conclude from Experiments 1 and 2 together that 15-month-olds have relatively immature generalization abilities in this context; they appear to link the efficacy of an action only to a specific kind of object. Adults clearly possess strong generalization capacities, particularly in the context of relating actions to effects (e.g., Kiesel & Hoffmann, 2004; Pfister, Kiesel, & Melcher, 2010). For instance, suppose you are in a new hotel and standing by an elevator on the ground floor. Next to the elevator, there is a button as well as a switch. Although you have never had experience with this particular elevator before, our intuition is that you would most likely press the button rather than flip the switch and expect that this pressing action would summon the elevator. In contrast, the results of Experiment 2 suggest that 15-month-olds would not do likewise unless the elevator and button

522

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

looked identical to one they had already experienced. This interpretation is consistent with prior research that infants’ learning and imitation may be quite ‘‘bound’’ to particular contexts (cf. Butler & Rovee-Collier, 1989; Rovee-Collier, Griesler, & Earley, 1985). Alternatively, we suggest that the results of Experiments 1 and 2 are consistent with infants quite rationally treating generalization as a problem of categorization. Given the high similarity of the two demonstration objects, the distinct test object in Experiment 2 would less likely be considered a member of the same category than the test object in Experiment 1, and thus generalization is less likely. Note that this perspective is consistent with our example of adults’ responses to the novel elevator; there, adults have at least one additional piece of information—knowledge that the object in front of them is an elevator—making generalization to the button likely. Similarly, although buttons on an MP3 player and a DVD player might look similar, only one will play a movie and adults do not press the MP3 button expecting a movie to ensue.1 Adults make these inferences because they know that the objects are members of the same category or different categories. Likewise, we argue that the more likely infants are to recognize the test object as a member of the same category as the demonstration objects, the more likely they will be to generalize. Accordingly, in Experiment 1, the objects were all highly similar and likely to be considered as members of a single category; hence, infants generalized to the test object. However, in Experiment 2, the test object was distinct from the two highly similar demonstration objects and, thus, was likely to be considered from a different category; hence, infants did not generalize to the test object. Experiment 3 tested this account based on interaction between generalization and categorization in a further way by increasing the variability of the demonstration objects. We showed 15-month-olds a different demonstration experience in which the effective and ineffective objects not only had different handles but also had different appearances altogether. Whereas children may have taken the similar demonstration toys in Experiments 1 and 2 to specify a narrow category, and hence a narrow generalization of the efficacy of actions, when presented with distinct demonstration toys, children might construe the objects more broadly as members of a single encompassing category and, hence, be likely to generalize the effective action again as in Experiment 1. Experiment 3 The third experiment retested the condition from Experiment 2 and contrasted it with a case in which the two demonstration objects likewise had different handles and efficacy but also had different overall appearances. Method Participants The final sample consisted of 48 full-term 15-month-olds (28 girls and 20 boys, mean age = 14.94 months, range = 14.00–16.00) who were recruited through collaboration with another infant laboratory. An additional 20 infants were tested but not included in the final sample because they failed to produce at least one of the demonstrated actions (n = 14), because they were fussy or uncooperative in general (n = 5), or because of experimenter error (n = 1). The sample was predominantly Caucasian, although other ethnicities and races were represented. Participants appeared to come from a range of middle-class families, although no particular socioeconomic status data were collected. Materials The same single-activator bunny and cow toys from Experiments 1 and 2 were used in Experiment 3 along with a new dual-handle test toy. This toy was made of green plastic and shaped like a pyramid (see Fig. 2). The base was a 17.5-cm square narrowing to a 10-cm square at the top, on which sat an orange dome 8 cm in diameter. The toy was 15.5 cm in height overall. Both the push-button and the pull-lever protruded approximately 3 cm out from the lower part of the toy; the push-button could be 1

We are grateful to Roland Pfister for pointing out this example.

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

523

Fig. 2. Novel test toy used in Experiment 3.

positioned to either side of the pull-lever in order to counterbalance the left/right positioning of the handles. As with the test toys in the prior experiments, neither handle of the pyramid toy elicited any response when manipulated. Two new warm-up toys were also used in Experiment 3: a pair of large plastic cars and a set of stackable multicolored rings. Procedure The warm-up phase was identical to that in Experiments 1 and 2 except that the different stimuli described above were used. For the demonstration trials, infants were randomly assigned to either the same or varied condition. The same condition was identical in procedure to that of Experiment 2. Children saw alternate trials involving an effective toy and an ineffective toy from the same set (either the bunny toys or the cow toys, counterbalanced). In the varied condition, infants were shown the effective object from one set and the ineffective object from the alternative set on the demonstration trials. Whether the first object was from the bunny or cow set and whether the first object was effective or ineffective were both counterbalanced across infants. Except for the nature of the toys in the varied condition, the demonstration trials proceeded as described for Experiments 1 and 2. After the demonstration trials, the new pyramid toy with both activators (the test toy) was brought out and placed in front of the infant with one activator on the infant’s left and the other on the right. The left/right position of the handle that had been effective on the demonstration trials was counterbalanced across infants. As in the prior experiments, no demonstration was provided with the test toy; the infant was simply allowed to play with it for 30 s. Coding The same coding system described for Experiments 1 and 2 was used. Trials were scored by two observers working together. Reliability was examined by having a third coder independently score 25% of the data (6 children in each condition). Overall agreement on the first touch and first manipulation behaviors for the test trial was 97%, and overall agreement on the durations of touching and manipulating during the demonstration and test phases was also 97% (kappa = .94). Results Demonstration phase The durations of time infants executed the target actions on the effective and ineffective handles in each condition of Experiment 3 are displayed in Table 1. Infants in the same condition spent on average 12.96 s manipulating the handle of the effective toy and 14.58 s manipulating the handle of the

524

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

ineffective toy; this is not a significant difference, t(23) = 0.56, p = .29. Considered individually, 11 infants spent more time manipulating the effective handle and 13 infants spent more time manipulating the ineffective handle, a distribution not different from that expected by chance, binomial test, p = .42. Similarly, infants in the varied condition spent on average 11.87 s manipulating the effective handle and 15.46 s manipulating the ineffective handle, again not a significant difference, t(23) = 1.05, p = .15. Here, 10 infants spent more time manipulating the effective handle, 12 spent more time manipulating the ineffective handle, and 2 spent the same amount of time manipulating the two handles; this distribution is not different from that expected by chance, binomial test, p = .42. Thus, as in Experiments 1 and 2, infants in both conditions of Experiment 3 had equivalent practice with the two target actions before they were presented with the test toy. Test trial The numbers of infants in each condition who first touched and first manipulated each of the handles on the test toy are shown in Table 2. Consistent with the results of Experiment 2, infants in the same condition did not initially choose one handle preferentially over the other. In this condition, 3 infants first touched both handles simultaneously and 1 infant did not touch either handle; these 4 infants were excluded from the analyses on first touch because their choice of handle was not clear. Of the remaining infants, 10 first touched the effective handle and 10 first touched the ineffective handle; this distribution is not different from chance. Similarly, 10 infants first manipulated the effective handle, 12 first manipulated the ineffective handle, and 2 did not manipulate either handle; this distribution is also not different from chance, binomial test, p = .42. In the varied condition, however, infants initially behaved like infants in Experiment 1. Here, 3 infants touched both handles simultaneously and were excluded, but of the remaining infants, 16 first touched the effective handle and only 5 first touched the ineffective handle. This distribution is significantly different from chance, binomial test, p = .01. Likewise, 17 infants first manipulated the effective handle and just 7 first manipulated the ineffective handle; this distribution is also significantly different from that expected by chance, binomial test, p = .03. We also directly compared infants’ initial responses in the two conditions of Experiment 3. An analysis on the first touch results showed that infants were more likely to choose the effective handle over the ineffective handle in the varied condition than in the same condition, v2(1, N = 41) = 3.03, p = .04, u = 0.27. A similar analysis on the first manipulation results likewise showed that infants were more likely to choose the effective handle over the ineffective handle in the varied condition than in the same condition, v2(1, N = 46) = 3.05, p = .04, u = 0.26. Thus, in Experiment 3, infants’ initial behavior on the test trial was different in the two conditions; infants in the varied condition systematically first touched and also first manipulated the effective handle rather than the ineffective handle, whereas infants in the same condition chose the two handles equally often. The mean lengths of time infants in each condition of Experiment 3 spent touching and also manipulating each of the two handles during the test trial are shown in Table 3. Infants touched the effective and ineffective handles for approximately the same durations in both the same and varied conditions. Similarly, they manipulated the two different handles for approximately the same durations in the each of the two conditions. For touching and also for manipulating, a 2 (Condition: same vs. varied)  2 (Handle: effective vs. ineffective) mixed ANOVA was conducted on the duration scores; these analyses yielded no significant main effects or interactions for either measure. Separate comparisons within each condition of the durations for touching the effective handle versus the ineffective handle and also for manipulating the effective handle versus the ineffective handle also did not indicate any significant differences. Thus, unlike Experiment 1, although infants in the varied condition acted on the effective handle first, there was no indication that infants in this condition (or in the same condition) persisted at interacting with one handle more than with the other. Discussion Like infants in Experiment 2, infants in the same condition in Experiment 3 showed no initial or sustained preference for the handle whose previous manipulation had led to an effect. However,

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

525

the results for the varied condition of Experiment 3 suggest that infants may generalize knowledge of an action’s efficacy to a novel object under certain circumstances. Infants in the varied condition experienced one action’s efficacy and another action’s inefficacy in the context of two distinct objects. When these infants were subsequently offered a third distinct object with handles affording both actions, they systematically chose to first touch and first manipulate the handle that had previously led to an effect. In this regard, infants in this condition acted similarly to those in Experiment 1, where infants displayed knowledge of an action’s efficacy by choosing the previously effective handle on a test object similar to the demonstration objects. Computational model Our hypothesis was that infants’ generalization of efficacy action is motivated by how they categorize the demonstration and test objects. To explore this hypothesis further, we constructed a computational model by extending the RMC (Anderson, 1990; Sanborn, Griffiths, & Navarro, 2006; Sanborn et al., 2010) to interpret these data. We suppose that as each toy was observed, infants assigned it to a category balancing two cognitive goals: (a) trying to assign objects to categories that already have many entities in them (i.e., have as few categories as possible) and (b) trying to assign objects such that features are homogeneous within categories. There are two parameters in the model: a, which captures the tendency for objects to be placed in the same category, and b, which captures how much the model prefers that categories have homogeneous features. In the full version of the RMC, the first object observed is placed in a category by default. The second object is then placed either in the same category as the first object or in a new category. Placing the second object in the initial category is more likely when the two demonstration objects are the same than when they are different; when they are the same, more features of the two objects match. That said, it is still possible to place the two demonstration objects in the same (broader) category when they differ in perceptual appearance. The possibility of this happening depends on how much stronger b is than a—that is, on how much the model cares about feature homogeneity as opposed to category size. Mathematical details for this assignment are provided in the Appendix. The hybrid test object is then treated in a similar way regarding its categorization, and inferences about generalization depend on whether the hybrid object is placed in the same category as both, one, or neither of the demonstration objects. When the two demonstration objects are categorized together, how should the hybrid object be categorized? If the two demonstration objects share a similar feature, then if the hybrid object also shares that feature, it is likely to be categorized with the demonstration objects. If the hybrid object does not share that feature, this categorization is less likely. Intuitively, when the first two objects share a common feature (i.e., in the same condition), the test object is unlikely to be placed in a category that is more homogeneous without its presence. But when the two demonstration objects do not share a common feature but are categorized together, the hybrid object might be more likely to be placed in that category given the preference to avoid creating new categories. We used a hierarchical learning framework (based on Kemp, Perfors, & Tenenbaum, 2007) that not only learns what categories objects are in but also can learn the values of parameters such as b and a and specifically had the model learn the value of b (see the Appendix for more details). When the demonstration objects are identical in appearance but differ in handle and efficacy (i.e., the same condition), the model should learn that b is strong. When the demonstration objects differ in appearance, handle, and efficacy (the varied condition), the model should learn that b is weak. To apply this model to the procedure used here, we used the RMC algorithm but automatically assigned the first two objects to the same category and then adjusted b depending on the homogeneity of the features of that category. Assigning the two demonstration objects to the same category stemmed from a pedagogical assumption we made about the demonstration—that the child believed the modeler was showing him or her these objects for a reason (i.e., to learn something about the demonstration). This assumption follows from both computational models of pedagogy (e.g., Shafto & Goodman, 2008) and beliefs infants have about others’ intentions (e.g., ‘‘natural pedagogy’’; see Csibra & Gergely, 2009; see also Gweon et al., 2010).

526

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

We then used a Monte Carlo simulation to collect 10,000 samples of how the (different hybrid) test object was categorized and measured the proportion of samples in which the model assigned the third object to the original demonstration category as opposed to a new one. When the two demonstration objects shared a common appearance, the different hybrid test object was also placed in that category 41% of the time. When the two demonstration objects did not share a common appearance, the test object was also placed in the initial category more frequently—52% of the time.2 These results at least qualitatively capture the difference between the same and varied conditions in Experiment 3, as infants generalized more robustly in the varied condition as compared with the same condition. When all three objects were the same in appearance (as in Experiment 1), the model put all three objects in the same category 80% of the time. In addition to predicting the difference between the two conditions of Experiment 3, the RMC we used also captured a difference between the results of Experiment 1 and those of Experiment 3 that we did not originally anticipate. In Experiment 1, along with choosing the effective handle first, children also spent more time overall manipulating the effective handle than the ineffective one on the test toy. This persistence difference was not found in Experiment 3. Recall, however, that the test toys in both experiments were ineffective. If we use the same RMC as before but with another feature representing the ‘‘failed’’ efficacy of the supposedly effective test handle—which becomes apparent only after the first touch and manipulation—the probability that all three objects are in the same category then drops. However, it does so less drastically when all three objects were the same in appearance as in Experiment 1 (80% to 62%) compared with when they were all different in appearance as in the varied condition of Experiment 3 (52% to 26%). Thus, when the model receives information that the test object’s supposedly effective handle is actually ineffective, it is more likely to continue to place the test object in the same category with the demonstration objects (and hence infants are more likely to continue to try to make it work) when the three objects are the same in appearance than when all three objects are different. Although this computational model does not mirror the exact probabilities with which children touched or manipulated the effective handle first (e.g., in the varied condition in Experiment 3, it predicts generalization only 52% of the time), a strength of this framework is that it captures all of the important differences among the conditions we presented, particularly in its ordinal predictions. Critically, in this model, we equated the relevance of the three features involved (appearance, handle, and efficacy). Manipulating the relative strength of these features (or adding features that independently code perceptual features such as shape and color) might allow the model to present a more precise quantitative fit of these data. Regardless of these remaining nuances, the model presented here at least provides a good first pass at describing how a rational approach might characterize infants’ generalization capacities in this context. General discussion The results of Experiments 1 to 3 suggest a multifaceted pattern of influences on infants’ generalization of goal-directed actions from imitation. In Experiment 1, although 15-month-olds willingly imitated both effective and ineffective actions demonstrated on a single type of object, they preferentially performed the effective action when offered a version of the object affording both actions. These results indicate that when infants learn from imitating others’ goal-directed actions, they are not just blindly mimicking the actions or simply acquiring information about how to move an object’s manipulandum. Instead, they are also sensitive to whether an action generates an effect beyond the handle’s simple movement; they learn the cause–effect relation involved. However, Experiment 2 demonstrated that 15-month-olds do not always generalize the efficacy of an action they learn from imitation. When given the same demonstrations as in Experiment 1 but then 2 The ordinal relation of these numbers was robust with respect to changes in a; when a was 0.5 (a stronger preference for fewer categories), the numbers were 35% and 66%, respectively, and when a was 2 (a weaker preference for fewer categories), the numbers were 28% and 36%, respectively. Even at extreme values of a such as 10 (which yielded 8% and 11%, respectively), the relation held. That is, for all values of a that we tried, the model made the same qualitative prediction; it predicted more generalization when the two demonstration objects did not share a common appearance.

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

527

offered a different kind of object affording both actions, infants chose randomly between the two handles and engaged in the two actions for equivalent amounts of time. Infants treated generalization as a problem of category membership; in Experiment 2, the test object was construed as from a new category, and hence infants did not generalize efficacy. In Experiment 3, we showed that with exposure to a broader sample of information during the demonstrations, infants did now generalize to a distinct test object. Thus, our key finding is not simply that infants do or do not generalize from imitating goal-directed behaviors; it is that infants’ generalization is a flexible process based on whether the object context stipulates object similarities or object distinctions. These differences across experiments are captured well by a computational framework based on a rational model of categorizing objects. This model makes a particular assumption about the pedagogical intent of the person demonstrating the efficacy, namely that the demonstration objects were not chosen randomly, but rather to convey information about category membership. The pedagogical assumption built into the model is that the two demonstration objects are members of the same category—that is, they represent an effective member and an ineffective member of that category—and the model makes appropriate qualitative predictions only given this assumption. We believe this assumption is justified by virtue of 15-month-olds’ ability to integrate the intentionality of an interactive partner with the sampling process and current data when making inferences (Gweon et al., 2010). That said, infants could have made other assumptions. For instance, infants could assume that the imitation partner is showing them two objects that are specifically members of different categories. Computationally, having our model make this assumption does not capture the results presented here.3 More important, such an assumption seems counterintuitive to the remainder of the demonstration. If the intention of the demonstrator was to show that these objects were members of different categories, it is likely that he or she would act quite differently toward them (e.g., separate them more linguistically, spatially, or in the reactions to their efficacy). This was not the case, and because infants are sensitive to the intentional actions of others (e.g., Carpenter, Nagell et al., 1998), it is unlikely that they made this assumption. Finally, it is important to point out that generalization differed in a subtle way between Experiment 1 and Experiment 3. Infants in the varied condition of Experiment 3 did not manipulate the effective handle for longer durations overall than the ineffective handle, although they did choose to manipulate it first. In contrast, in Experiment 1, infants strongly persisted in manipulating the previously effective handle, even though it did not elicit an effect on the test object. The difference in strength between the results of Experiment 1 and those of the varied condition of Experiment 3 is consistent with developmental research suggesting that generalizing from imitation is more fragile or demanding than imitating in and of itself (e.g., Barnat, Klein, & Meltzoff, 1996; Bauer & Dow, 1994; Hayne, MacDonald, & Barr, 1997). Like the various within-experiment results, this between-experiment difference is also qualitatively captured by the computational framework we presented given the particular pedagogical assumption.

Categorization versus encoding We have suggested that for 15-month-olds shared category membership leads to the generalization of efficacy, and we have proposed a framework for describing how that generalization might occur. In this way, generalization is a cognitive process based on categorical information. Our approach is consistent with theories of infant categorization that focus on parts and their configurations (e.g., Rakison & Butterworth, 1998) because in order to categorize the objects together, infants must attend to the relation between the objects’ parts (in particular the handles) as well as to the other perceptual features of the objects. Another way of considering these results is in terms of generalization being a process of action. For instance, proponents of a ‘‘theory of event coding’’ (e.g., Hommel, Müsseler, Aschersleben, & Prinz, 2001; Kunde, 2001) would suggest that what infants represent is the action they perform, and the effects of generalization observed here are not a categorical process but rather relate to encoding spec3

This version of the model predicts no differences among the three experiments or between the two conditions in Experiment 3.

528

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

ificity. Thus, an action’s efficacy is encoded in the context of the object, and the demonstration phase sets that context. In Experiments 1 and 2 and in the same condition of Experiment 3, the context is specific to a particular kind of object, whereas in the varied condition of Experiment 3, the context is broader. This approach is consistent with our findings as well as with a set of developmental phenomena (albeit on older children, see Karbach, Kray, & Hommel, 2011; Kray, Eenshuistra, Kerstner, Weidema, & Hommel, 2006). In this view, the computational model we presented could be conceptualized as a way of describing how infants differentially organize the various contexts across experiments and conditions. The current data and model cannot determine between these objectcentered and action-centered theoretical possibilities—that is, whether the effect of context is present in the representation of the action the infant engages in or in the way the infant chooses to categorize the various objects to which the action may be applied. Teasing apart these two potential approaches to the effect of context on generalization could be a focus for future research. Extensions to the current work and conclusion Our computational model suggests that the problem of generalization may be more about learning the conceptual structure that exists among the relevant set of objects. To this end, we could imagine extending these experiments in various ways. For instance, we would predict that presenting children with additional common or distinctive features, such as giving all of the objects the same label versus giving the effective and ineffective objects different labels during the demonstration, might change how infants generalize. Similarly, several researchers have shown the importance of discovering efficacy in causal learning (e.g., Kushnir & Gopnik, 2005; Sobel & Sommerville, 2010). These data suggest that the amount of generalization might increase if children acted first on these toys themselves as opposed to imitating another’s actions; conversely, generalization might decrease if children only observed another act on them (i.e., are not allowed to produce the actions themselves on the demonstration toys). If this latter finding held, it would potentially be more consistent with infants’ encoding context in their representation of the action than with our notion that context is integrated into the process of object categorization. At this point, however, we are agnostic to this claim. To conclude, the current experiments suggest that 15-month-olds distinguish between cases where an action’s efficacy seems specific to a particular object type and cases where the efficacy pertains to a broader class of objects or possibly to the action itself. Thus, when imitating goal-directed actions, infants are not just replicating actions, learning affordances, or perceiving effects; instead, they are integrating these aspects of the situation together with one another and with other relevant information in the situation. This integration sometimes leads to tightly linking an action’s efficacy only to the object on which it was observed and practiced. However, in other contexts, the integration leads to a looser inference regarding the action’s efficacy and, thus, generalizes it to a range of novel objects. Thus, infants’ generalization from imitating goal-directed behaviors is a flexible and rational process dependent on their own evaluation of newly confronted objects and their category membership. Acknowledgments Experiments 1 and 2 were part of the research submitted by the first author to Tufts University for her master’s degree. Experiment 3 was supported by National Science Foundation (NSF) Grant DLS0518161 to D.M.S. We thank all of the parents and children who participated in this research. We also thank Esra Aksu, Claire Cook, Jessica Fradkin, Daniel Greenwald, Everett Hendler, Emily Hopkins, Sarah Kerstein, Alice McMahon, Corrine Mahoney, Mara Sacks, Rachel Shelley-Abrahamson, Kristen Sylvester, and Johanna Thompson-Holland for helping with data collection and analysis. Appendix A Mathematical details for computational model (initially described in Sobel, Buchanan, Butterfield, & Jenkins, 2010) We used the fully rational Dirichlet process as described by Sanborn and colleagues (2006) and also incorporated hierarchical learning of b. Our comparison models were modified

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

529

versions of this model. We first describe the algorithms that each model used to sample the category assignments and then specify our exact equations. The fully rational model used Gibbs sampling. The objects were initialized to random clusters. At each step, we resampled b taking the cluster assignments as given and then resampled each cluster assignment with this value of b given the other cluster assignments. b was sampled using an independence chain Metropolis–Hastings sampler; new values of b were proposed from Exponential (1), and then accepted with a probability proportional to the likelihood ratios of the currently proposed value of b given the category assignments. That is, if the ratio was greater than 1, the proposed new value was accepted, and if it was less than 1, the new value was accepted if a randomly chosen number between 0 and 1 was less than the ratio. Once we had iterated the Metropolis–Hastings component 10 times, we proceeded with the rest of the Gibbs sampling step. We repeated the Gibbs sampling until we had 10,000 samples, sampling the table assignments every 10th step. This model did not fit the data well; the three objects were categorized as members of the same category 27% of the time when the first two objects shared features and the third was different, but only 20% of the time when all three objects had different features. The other two models were modified versions of this sampler in which the first two objects were always assigned to the first cluster and their cluster assignments were never resampled. The only values sampled were b (which was only learned from the cluster assignments of the first two objects) and the cluster assignment of the test object. We repeated this 10,000 times and counted the proportion of times that the objects were all categorized together. One of these versions was also without hierarchical learning; this was the same except that we did not sample b (it was always set at 1). This model also did not fit the data well, showing no difference between cases in which the two demonstration objects shared a feature or were different; the test object was placed in the same category as the demonstration objects 55% of the time in both conditions. Fits for the hierarchical learning model are provided in the text. To model additional feature data, we added a feature that represented the efficacy of the previously effective handle. We now describe the equations in the samplers. To perform the Gibbs sampling, we needed to know the probability that an object was a member of a cluster given the parameters, features, and what objects were assigned to what clusters. The RMC assumes that features are independent given the category. We compared the likelihood of each cluster assignment using the following equation:

PðC i ¼ cjCi ; F; a; bÞ / PðC i ¼ cjC i ; aÞPðF i ¼ f jC i ¼ c; bÞ; where F is a vector of features, C is a vector of cluster assignments, and a and b are parameters. We set a = 1 In this equation, and all of what follows, Ci refers to all members of C except Ci whereas Ci refers to all Ck such that k < i. The categorization process was a Dirichlet process that assigned a probability to each cluster as follows:

PðC i ¼ cjC i ; aÞ ¼

8 < #fj:C j Ci ;C j ¼cg ; if c is an exsisting category #fj:C j C i gþa :

a

#fj:C j C i gþa

; if c is a new category

;

where Ci is the vector of category assignments of all previous objects, #fj : C j C i g counts the number of previously categorized objects, and #fj : C j C i ; C j ¼ cg counts the number of previously categorized objects that were placed in category c. The features were assumed to come from a multinomial with a Dirichlet prior, with nf possible states that were always equally likely:

PðF i ¼ f jFi ; C i ¼ c; bÞ ¼

#fj :g þ b : #fj : C j ¼ cg þ nf b

For the initial model, there was only one feature with nf = 3 states corresponding to the three shapes. In one version of the model, we added a binary feature ðnf ¼ 2Þ corresponding to the efficacy of one of the handles.

530

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

The model also learned b by comparing the likelihood ratios of two candidate values. The likelihoods were proportional to:

 #fj : C c ; F ¼ f g þ x #fj:C ¼cg j i j ; P b ¼ xjC; F; aÞ / ex Gc Gf Gi¼1 j #fj : C j ci g þ nf x where ci was the set of previous members of that category. Thus, a likelihood was proportional to the product of the probabilities of generating each feature sequentially using b for each feature and category. References Anderson, J. R. (1990). The adaptive character of thought. Hillsdale, NJ: Lawrence Erlbaum. Barnat, S. B., Klein, P. J., & Meltzoff, A. N. (1996). Deferred imitation across changes in context and object: Memory and generalization in 14-month-old infants. Infant Behavior and Development, 19, 241–251. Bauer, P. J. (1992). Holding it all together: How enabling relations facilitate young children’s event recall. Cognitive Development, 7, 1–28. Bauer, P. J., & Dow, G. A. (1994). Episodic memory in 16- and 20-month-old children: Specifics are generalized but not forgotten. Developmental Psychology, 30, 403–417. Bauer, P. J., & Hertsgaard, L. A. (1993). Increasing steps in recall of events: Factors facilitating immediate and long-term memory in 13.5- and 16.5-month-old children. Child Development, 64, 1204–1223. Bonawitz, E., Shafto, P., Gweon, H., Goodman, N. D., Spelke, E., & Schulz, L. E. (2011). The double-edged sword of pedagogy: Instruction limits spontaneous exploration and discovery. Cognition, 120, 322–330. Boyd, R., Richerson, P. J., & Henrich, J. (2011). The cultural niche: Why social learning is essential for human adaptation. Proceedings of the National Academy of Sciences (USA), 108, 10918–10925. Brugger, A., Lariviere, L. A., Mumme, D. L., & Bushnell, E. W. (2007). Doing the right thing: Infants’ selection of actions to imitate from observed event sequences. Child Development, 78, 806–824. Bushnell, E. W., & Boudreau, J. (1998). Exploring and exploiting objects with the hands during infancy. In K. J. Connolly (Ed.), The psychobiology of the hand: Clinics in developmental medicine (pp. 144–161). London: MacKeith. McDonald Institute Monograph Series. Bushnell, E. W., Sidman, J., & Brugger, A. E. (2006). Transfer according to the means in human infants: The secret to generative tool-use. In V. Roux & B. Brill (Eds.), Stone knapping: The necessary conditions for a uniquely hominid behaviour (pp. 303–317). Cambridge, UK: McDonald Institute for Archaeological Research. Butler, J., & Rovee-Collier, C. (1989). Contextual gating of memory retrieval. Developmental Psychobiology, 22, 533–552. Carpenter, M., Akhtar, N., & Tomasello, M. (1998). Fourteen- through 18-month-old infants differentially imitate intentional and accidental actions. Infant Behavior & Development, 21, 315–330. Carpenter, M., Nagell, K., & Tomasello, M. (1998). Social cognition, joint attention, and communicative competence from 9 to 15 months of age. Monographs of the Society for Research in Child Development, 63(4, Serial no. 255). Chen, Z., & Siegler, R. S. (2000). Across the great divide: Bridging the gap between understanding of toddlers’ and older children’s thinking. Monographs of the Society for Research in Child Development, 65, 1–105. Cohen, L. B., & Strauss, M. S. (1979). Concept acquisition in the human infant. Child Development, 50, 419–424. Csibra, G., & Gergely, G. (2009). Natural pedagogy. Trends in Cognitive Science, 13, 148–153. Elsner, B., & Aschersleben, G. (2003). Do I get what you get? Learning about the effects of self-performed and observed actions in infancy. Consciousness and Cognition, 12, 732–751. Elsner, B., & Pauen, S. (2007). Social learning of artifact function in 12- and 15-month-olds. European Journal of Developmental Psychology, 4, 80–99. Gergely, G., Bekkering, H., & Király, I. (2002). Rational imitation in preverbal infants. Nature, 415, 755–756. Gopnik, A., Glymour, C., Sobel, D. M., Schulz, L. E., Kushnir, T., & Danks, D. (2004). A theory of causal learning in children: Causal maps and Bayes nets. Psychological Review, 111, 3–32. Gweon, H., Tenenbaum, J. B., & Schulz, L. E. (2010). Infants consider both the sample and the sampling process in inductive generalization. Proceedings of the National Academy of Sciences, 107, 9066–9071. Harris, P. L., & Koenig, M. A. (2006). Trust in testimony: How children learn about science and religion. Child Development, 77, 505–524. Hayne, H., MacDonald, S., & Barr, R. (1997). Developmental changes in the specificity of memory over the second year of life. Infant Behavior and Development, 20, 233–245. Herbert, J., Gross, J., & Hayne, H. (2006). Age-related changes in deferred imitation between 6 and 9 months of age. Infant Behavior and Development, 29, 136–139. Herbert, J., & Hayne, H. (2000). Memory retrieval by 18–30-month-olds: Age-related changes in representational flexibility. Developmental Psychology, 36, 473–484. Hommel, B., Müsseler, J., Aschersleben, G., & Prinz, W. (2001). The theory of event coding: A framework for perception and action planning. Behavioral and Brain Sciences, 24, 869–878. Karbach, J., Kray, J., & Hommel, B. (2011). Action–effect learning in early childhood: Does language matter? Psychological Research, 75, 334–340. Kemp, C., Perfors, A., & Tenenbaum, J. B. (2007). Learning hypotheses with hierarchical Bayesian models. Developmental Science, 10, 307–321. Kiesel, A., & Hoffman, J. (2004). Variable action effects: Response control by content-specific effect anticipation. Psychological Research,68, 155–162.

D.J. Yang et al. / Journal of Experimental Child Psychology 116 (2013) 510–531

531

Kray, J., Eenshuistra, R., Kerstner, H., Weidema, M., & Hommel, B. (2006). Language and action control: The acquisition of action goals in early childhood. Psychological Science, 17, 737–741. Kunde, W. (2001). Response–effect compatibility in manual choice reaction tasks. Journal of Experimental Psychology: Human Perception & Performance, 27, 387–394. Kushnir, T., & Gopnik, A. (2005). Young children infer causal strength from probabilities and interventions. Psychological Science, 16, 678–683. Liu, J., Golinkoff, R. M., & Sak, K. (2001). One cow does not an animal make: Young children can extend novel words at the superordinate level. Child Development, 72, 1674–1694. Lyons, D. E., Young, A. G., & Keil, F. C. (2007). The hidden structure of overimitation. Proceedings of the National Academy of Sciences of the United States of America, 104, 19751–19756. Meltzoff, A. N. (1988a). Infant imitation and memory: Nine-month-olds in immediate and deferred tests. Child Development, 59, 217–225. Meltzoff, A. N. (1988b). Infant imitation after a 1-week delay: Long-term memory for novel acts and multiple stimuli. Developmental Psychology, 24, 470–476. Meltzoff, A. N. (1995). Understanding the intentions of others: Re-enactment of intended acts by 18-month-old children. Developmental Psychology, 3, 838–850. Meltzoff, A. N., & Moore, M. (1977). Imitation of facial and manual gestures by human neonates. Science, 198, 75–78. Meltzoff, A. N., & Moore, M. (1983). Newborn infants imitate adult facial gestures. Child Development, 54, 702–709. Meltzoff, A. N., Waisman, A., & Gopnik, A. (2012). Learning about causes from people: Observational learning in 24-month-old infants. Developmental Psychology, 48, 1215–1228. Nagell, K., Olguin, R. S., & Tomasello, M. (1993). Processes of social learning in the tool use of chimpanzees (Pan troglodytes) and human children (Homo sapiens). Journal of Comparative Psychology, 107, 174–186. Oakes, L. M., Madole, K. L., & Cohen, L. B. (1991). Object examining: Habituation and categorization. Cognitive Development, 6, 377–392. Pfister, R., Kiesel, A., & Melcher, T. (2010). Adaptive control of ideomotor effect anticipations. Acta Psychologica, 135, 316–322. Quinn, P. C., & Bhatt, R. S. (2005). Learning perceptual organization in infancy. Psychological Science, 16, 511–515. Rakison, D. H., & Butterworth, G. (1998). Infant attention to object structure in early categorization. Developmental Psychology, 34, 1310–1325. Rovee-Collier, C., Griesler, P. C., & Earley, L. A. (1985). Contextual determinants of retrieval in three-month-old infants. Learning & Motivation, 16, 139–157. Sanborn, A. N., Griffiths, T. L., & Navarro, D. J. (2010). Rational approximations to rational models: Alternative algorithms for category learning. Psychological Review, 117, 1144–1167. Sanborn, A. N., Griffiths, T. L., & Navarro, D. J. (2006). A more rational model of categorization. In R. Sun & N. Miyake (Eds.), Proceedings of the 28th annual conference of the Cognitive Science Society. Austin, TX: Cognitive Science Society. Schulz, L. E., Hooppell, K., & Jenkins, A. (2008). Judicious imitation: Young children imitate deterministic actions exactly, stochastic actions more variably. Child Development, 79, 395–410. Shafto, P., & Goodman, N. (2008). Teaching games: Statistical sampling assumptions for learning in pedagogical situations. In B. C. Love, K. McRae, & V. M. Sloutsky (Eds.), Proceedings of the 30th annual conference of the Cognitive Science Society (pp. 1632–1637). Cognitive Science Society: Austin, TX. Sobel, D. M., Buchanan, D. W., Butterfield, J., & Jenkins, O. C. (2010). Interactions between causal models, theories, and social cognitive development. Neural Networks, 23, 1060–1071. Sobel, D. M., & Sommerville, J. A. (2009). Rationales in children’s causal learning from others’ actions. Cognitive Development, 24, 70–79. Sobel, D. M., & Sommerville, J. A. (2010). The importance of discovery in children’s causal learning from interventions. Frontiers in Developmental Psychology, 1, 176–183. Tomasello, M. (1999). The cultural origins of human cognition. Cambridge, MA: Harvard University Press. Trauble, B., & Pauen, S. (2007). The role of functional information for infant categorization. Cognition, 105, 362–379. Want, S. C., & Harris, P. L. (2002). How do children ape? Applying concepts from the study of non-human primates to the developmental study of ‘‘imitation’’ in children. Developmental Science, 5, 1–41. Yang, D., Sidman, J., & Bushnell, E. W. (2010). Beyond the information given: Infants’ transfer of actions learned through imitation. Journal of Experimental Child Psychology, 106, 62–81. Younger, B. A., & Fearing, D. D. (2000). A global-to-basic level trend in early categorization: Evidence from a dual-category habituation task. Infancy, 1, 47–58.