Conceptually Rich, Perceptually Sparse Object Representations

1 Conceptually rich, perceptually sparse object representations Conceptually rich, perceptually sparse object representations 2 Abstract Six-month-old infants can store representations of multiple objects in working memory, but do not always remember the objects’ features (e.g. shape). Here, we asked whether infants’ object representations (1) may contain conceptual content, and (2) may contain this content even if perceptual features are forgotten. We hid two conceptually distinct objects (a humanlike doll and a nonhuman ball) one at a time in two separate locations, and then tested infants’ memory for the first-hidden object by revealing either the original hidden object or an unexpected other object. Using looking time, we found that infants remembered the categorical identity of this object, but failed to remember its perceptual identity. Together, our results suggest that young infants may encode conceptual category in a representation of an occluded object, and may do so even when perceptual features are lost. Conceptually rich, perceptually sparse object representations 3 Introduction What information is encoded in an object representation stored in working memory? Maintaining the surface features (such as color, shape, or orientation) of items in working memory is known to be costly: as the number of to-be-remembered items increases, fewer features are retained (Kibbe & Feigenson, 2016; Kibbe & Leslie, 2013; Pylyshyn, 2004; Zosh & Feigenson, 2012; see Kibbe, 2015, for a review). For example, 6-month-olds who observe two featurally distinct objects hidden sequentially in two separate locations will remember the featural identity of only the last object they saw hidden (e.g. Káldy & Leslie, 2005; Kibbe & Leslie, 2016). However, infants are surprised when the forgotten first-hidden object is revealed to have vanished completely, suggesting they remembered its existence (Kibbe & Leslie, 2011). Evidence from older infants suggests that they may use categorical information about objects to individuate those objects. Xu and Carey (1996) found that, before 12 months, infants were unable to decide, for example, that a duck pulled from behind a screen and placed back again was distinct from a truck that is subsequently pulled from behind the same screen, even though the objects were featurally distinct. However, by 12 months, infants could use their knowledge of the objects’ categories to individuate these objects (Xu, Carey, Quint, 2004; Xu, Carey, & Welch, 1999). Testing 10-montholds, Bonatti and colleagues found that if one of the two objects was a doll’s head, infants individuated the objects, but if the objects consisted of two featurally distinct doll’s heads, infants failed (Bonatti, Frot, & Mehler, 2002; see also Bonatti, Frot, Zangl, & Mehler, 2005). Surian and Caldi (2010) found that when one of two featurally distinct objects was animate and the other was inanimate, 10-month-olds individuated the Conceptually rich, perceptually sparse object representations 4 objects; however, infants failed when both objects were animate. Together, these studies suggest that the category HUMAN or ANIMATE may be used for individuation by younger infants when surface features are not (for evidence of a privileged face category in adults see e.g. Liu, Harris, & Kanwisher, 2009; Peelen & Caramazza, 2012; Stein, Sterzer, Peelen, 2012). The above studies, however, examined only individuation. Leslie, Xu, Tremoulet and Scholl (1998; Scholl & Leslie, 1999) drew a distinction between individuation and object identification, arguing that information that creates an object representation may not necessarily be bound to that representation. Infants may then know how many objects there are in a scene but have limited or no expectations regarding what properties they each have. Given that category information, when available, is more readily used to individuate objects than objects’ surface features, perhaps it may be encoded in an object representation — even when perceptual features are not. Indeed, Baillargeon and her colleagues have suggested that categorical information may hold a privileged position in infants’ representations of objects in physical events, including occlusion and containment (Baillargeon et al., 2009; Baillargeon et al., 2012). However, no work to date has examined whether categories that are used to individuate objects may also be encoded in an object representation, nor whether an object’s category may be encoded even when its surface features are not. Here, we directly tested the hypothesis that the conceptual content of an object representation in working memory may be richer than the perceptual content. By conceptual content, we mean the representation of a category, already stored in long- Conceptually rich, perceptually sparse object representations 5 term memory, which is accessed and applied top-down. By perceptual content, we mean features (e.g. shape, surface reflectance (color), spatial frequency) that are currently available bottom-up (Kanwisher, 2000). We used a two-object, two-location hiding method, which can measure both individuation and identification in 6-month-olds (Káldy & Leslie, 2005; Kibbe and Leslie, 2011, 2016). We tested 6-month-olds in particular because they can remember the existence, but not the featural identity, of the first-hidden of the two objects (Kibbe & Leslie, 2011), allowing us to explicitly test whether conceptual content may be encoded where featural content typically is not. In Experiment 1, we hid objects belonging to two categories 6-month-olds are known to discriminate: human and non-human object (e.g., Southgate et al. 2008) and probed infant’s memory for the first-hidden object by revealing either the original hidden object (control condition) or the unexpected other object (swap condition; between-subjects design). We reasoned that if conceptual information is encoded more readily than perceptual information in the object representation, infants should remember the categorical identity of the probed object. In Experiment 2, we ruled out the alternative perceptual-featural explanation for infants’ performance in Experiment 1. Experiment 3 replicated key conditions. Finally, in Experiment 4, we asked whether the categorical distinction between the objects improved infants’ memory for the featural identity of the probed object. Experiment 1: Human versus Object, Between-Category Change Participants Twenty-four healthy, full-term infants participated (mean age: 6 months 2 days; range: 5 months 18 days to 7 months 0 days; 13 girls). Infants were randomly assigned to Conceptually rich, perceptually sparse object representations 6 either the control condition (n = 12) or the swap condition (n = 12; between subjects, design). Sample size (n = 12 infants per condition) was determined prior to data collection based on the observed effect size (d = 1.13) in our previous study (Kibbe & Leslie, 2011), which used the same method as in the present study to examine infants’ looking times to structural changes versus no changes to the first-hidden object (1-β> .8, α=.05). An additional seven infants were tested but excluded from analysis due to fussiness (5) or equipment malfunction (2). Materials Stimuli consisted of two objects: a yellow and red striped ball (~9 cm in diameter) glued onto a 4.5 x 4.5 black foam-core platform, and a doll’s head with brown skin and eyes (~9 cm in diameter). A circular rigid blue fabric “bib” (11.5 cm diameter) was glued around the doll’s neck (see Figure 1). Two black foam-core occluders (28.5 x 17.75 cm) were used during the test trials. Stimuli were presented on a black wooden stage (130 x 43 x 50 cm). A curtain hid the stage area between trials. The experimenter wore long white gloves throughout the experiment and jingle bells around her right wrist. Procedure Infants were seated on a caregiver’s lap about 70 cm from the stage. At the beginning of the experiment, the experimenter lowered the curtain revealing the empty stage. She then drew infants’ attention to the front and back right, left, and center, and to the middle of the stage by jingling the bells she wore around her wrist, so that an observer who was watching the infant on a monitor could get a sense of the infant’s eye positions relative to the stage. The experimenter then raised the curtain, hiding the stage area. Conceptually rich, perceptually sparse object representations 7 Familiarization Trials. At the beginning of each familiarization trial, the experimenter placed the doll and the ball sequentially on the center of the empty stage, 7 cm apart. After 4 s, the experimenter moved the objects sequentially to the back of the stage, 55 cm apart, in the same order in which they were originally presented. After 8 s, the experimenter raised the curtain, hiding the stage. The experimenter repeated this procedure three more times for a total of four familiarization trials. The order and side of placement of the objects was counterbalanced across trials so that infants could not form any long-term associations between particular objects and locations. Order and side of placement of the objects on each trial was also counterbalanced across infants. Figure 1, left panel, shows two sample familiarization trials (see also Supplementary Video S1 for an example of a familiarization trial). There was no occlusion of objects during familiarization trials. Test Trials. At the beginning of each test trial, the experimenter placed the two occluders on the empty stage, 48 cm apart. The experimenter then placed the ball and the doll sequentially on the center of the stage, 7 cm apart, as in the familiarization trials. After 4 s, she then moved the objects behind the occluders in the order in which they were originally presented. After hiding the second object, the experimenter lifted the occluder hiding the first-hidden (harder-to-remember) object to reveal one of two possible outcomes: half of the infants saw the object that had been hidden there originally (e.g. ball hidden, ball revealed or doll hidden, doll revealed; control condition); the other half of infants saw that the object had been surreptitiously swapped for the other object (e.g. ball hidden, doll revealed or doll hidden, ball revealed; swap condition). Conceptually rich, perceptually sparse object representations 8 Infants saw three more test trials, for a total of four. The order of placement of the objects and their locations relative to each other across trials was the same as in the familiarization trials. Thus, on half the trials, the ball was hidden first, and on the other half the doll was hidden first. As in the familiarization trials, order and side of placement of objects on each trial was also counterbalanced across infants. Figure 1, right panel, shows a sample test trial for Experiment 1 (see also Supplementary Videos S2 and S3 for examples of control and swap test trials, respectively). Figure 1. Example familiarization trials (left) and test trial (right) for Experiment 1. During familiarization trials, infants observed the objects placed one at a time on the stage and then moved to the back of the stage. During test trials, infants observed the Conceptually rich, perceptually sparse object representations 9 objects placed on the stage and then hidden alternately behind occluders. The occluder hiding the first-hidden object was then removed to reveal either the object that had been hidden in that location originally (control outcome) or the other object (swap outcome). Order of presentation and alternating side of placement of objects was counterbalanced across trials during both familiarization and test. An observer who was naïve to condition (control or swap) and who was unable to see what infants were seeing measured infants’ looking duration following the test outcomes. When infants looked away from the stage for two consecutive seconds, the observer signaled the experimenter to terminate the trial. The experimenter then raised the curtain, hiding the stage. Infants’ looking times were later recoded offline, frame-byframe, by an observer who was naïve to condition, using Preferential Looking Coder (Libertus, 2011), and these looking times were analyzed. A second naïve observer also coded a random 50% of infants. Inter-observer agreement was high (mean r = .98). Results To correct for right skew, which is common for looking time data, all raw looking times were log transformed, and analyses were conducted on the log-transformed data (see Csibra et al., 2015; Hays, 1994). For the following planned analyses, we used traditional null hypothesis significance testing as well as Bayes Factor analysis using the Jeffrey-Zellner-Siow prior (Gallistel, 2009; Rouder et al., 2009) in order to quantify the evidence for or against the null hypothesis that infants were unable to remember the firsthidden object. A repeated-measures ANOVA with Trial (1st, 2nd, 3rd, or 4th) as a within-subjects factor and Condition (control or swap) as a between-subjects factor revealed no main effect of Trial (F(3,66) < 1, p = .76, ηp2 = .018) and no Trial X Condition interaction (F(3,66) < 1, p = .46, ηp2 = .038). We observed a significant main effect of Condition Conceptually rich, perceptually sparse object representations 10 (F(1,22) = 5.39, p = .03, ηp2 = .20), although Bayes factor analysis suggested anecdotal support for the alternative hypothesis that infants who saw the swap outcome looked longer than infants who saw the control outcome (JZS BF10 = 2.32). Figure 2 shows the mean raw and log looking times from Experiment 1. In Figure 3, we plot the cumulative probability of the individual mean log looking times from each infant in Experiment 1, providing a visualization of the degree of overlap of the distributions of the data from each condition. Greater overlap suggests that the data from each condition were generated by the same process, suggesting support for H0, while less overlap suggests that the data from each condition were generated by different processes, suggesting support for H1 (a relationship that is quantified by Bayes factor analysis). While we did not observe a significant interaction between Trial and Condition, previous studies using violation of expectation typically observe the largest effects on the first test trial (see e.g. Kibbe & Leslie, 2011 for an example using a similar method; see Figure S1 for trial-by-trial data; for additional information see Supplementary Materials). Analysis of infants’ log looking times on the first trial only showed that infants looked significantly longer at the swap outcome versus the control outcome (t(22) = 3.08, p=.006), with Bayes factor analysis suggesting substantial support for the alternative hypothesis (JZS BF10 = 8.039). Conceptually rich, perceptually sparse object representations 11 Figure 2. Mean looking times (top panel) and mean log looking times (bottom panel) for Experiments 1-4. Error bars show +/-1 SEM. Conceptually rich, perceptually sparse object representations 12 Figure 3. Cumulative probabilities of the log-scaled mean looking time data (in seconds) from Experiments 1-4. Conceptually rich, perceptually sparse object representations 13 Finally, because infants have shown a preference to look at faces over non-faces (e.g. Goren et al., 1975; Farroni et al., 2005), we asked whether the main effect of Condition may have been driven by a preference to look at dolls versus balls. We first averaged infants’ log looking times across test trials in which the ball was revealed and test trials in which the doll was revealed. We then conducted a repeated-measures ANOVA with Object Revealed (ball or doll) as a within-subjects factor and Condition (control or swap) as a between-subjects factor. There was a significant main effect of Object Revealed (F(1,22) = 4.67, p = .04, ηp2 = .18): infants looked slightly longer overall at the doll (raw Mcontrol = 5.65 s; raw Mswap = 7.95 s) versus the ball (raw Mcontrol = 4.08 s; raw Mswap = 6.11 s), with Bayes factor analysis suggesting anecdotal evidence for the alternative (JZS BF10 = 1.61). Critically, there was no Object Revealed X Condition interaction (F(1,22) < 1, p = .67, ηp2 = .008), suggesting that infants’ pattern of longer looking at the doll was not different across conditions. Discussion Previous work showed that infants consistently fail to remember the identity of the first-hidden of two objects that contrast in featural identity (e.g. triangle and disk; Kibbe & Leslie, 2011, 2016; Káldy & Leslie, 2005). By contrast, infants in Experiment 1 looked longer at the swap outcome when the two objects contrasted in categorical identity, suggesting they remembered the first-hidden object’s identity. However, it is important to rule out perceptual explanations for our results. The doll and the ball, while categorically distinct, were also highly distinct perceptually. Infants in Experiment 1 may have looked longer not because they encoded categorical information about the objects, Conceptually rich, perceptually sparse object representations 14 but because they detected a change to one or more of the many featural differences in the stimuli. To test this possibility, in Experiment 2 we kept the features of the objects constant while removing the categorical distinctions between the objects. Infants again were shown the ball and the doll, but this time the doll was inverted. Inversion is known to disrupt face processing in both infants (Farroni et al., 2005; Southgate et al., 2008) and adults (Stein, Sterzer, & Peelen, 2012), and infants regard inverted faces as categorically distinct from upright faces (e.g. Bonatti et al., 2002, 2005). If infants succeeded in Experiment 1 because of perceptual differences in the objects, they should succeed again in Experiment 2. However, if infants in Experiment 1 succeeded because they were processing the objects as categorically distinct, infants should fail. Experiment 2 also addresses the alternative possibility that infants may have succeeded because they remembered the identity of the last-hidden object and noticed that it (unexpectedly) appeared in the first-hidden location (though previous work suggests this is unlikely; Kibbe & Leslie, 2013). Experiment 2: Inverted Human versus Object, Perceptual Control Participants A separate group of 24 healthy, full-term infants participated (mean age: 6 months 14 days; range: 5 months 15 days to 7 months 0 days; 13 girls). Infants were randomly assigned to either the control condition (n = 12) or the swap condition (n = 12; between subjects design). An additional two infants were tested but excluded from analysis due to fussiness. Materials and Procedure Conceptually rich, perceptually sparse object representations 15 Stimuli and procedure were identical to Experiment 1, except that the doll’s head was inverted throughout the experiment. Figure 4 shows two sample familiarization trials (left) an example test trial (right). See also Supplementary Video S4 for an example familiarization trial, and Supplementary Videos S5 and S6 for example control and swap outcome test trials, respectively. Figure 4. Example familiarization trials (left panel) and test trial (right panel) from Experiment 2. Experiment 2 proceeded as in Experiment 1, except the doll was inverted throughout. Order of presentation and side of placement of objects was counterbalanced across trials during both familiarization and test. Conceptually rich, perceptually sparse object representations 16 A naïve observer, who was unable to see what infants were seeing, measured looking time during test trials. Infants’ looking times were later recoded offline by an observer who was naïve to condition, and these looking times were analyzed. A second naïve observer also coded a random 50% of infants. Inter-observer agreement was high (mean r = .96). Results As in Experiment 1, data were log-transformed to correct for right skew. A repeated-measures ANOVA with Trial (1st, 2nd, 3rd, or 4th) as a within-subjects factor and Condition (control or swap) as a between-subjects factor revealed no main effect of Trial (F(3,66) < 1, p = .60, ηp2 = .03) and no Trial X Condition interaction (F(3,66) < 1, p = .81, ηp2 = .01). There was also no main effect of Condition (F(1,22) < 1, p = .52, ηp2 = .020), suggesting infants looked equally at control and swap outcomes, although Bayes factor analysis suggested anecdotal support for the null (BF01 = 2.79; Figures 2 and 3). Analysis of infants’ looking times on the first trial revealed a similar pattern (t(22) = .689, p = .498, ; BF01 = 2.25; Figure S2). As in Experiment 1, we asked whether infants’ looking may have been driven by a preference for one object versus the other. We predicted that, unlike Experiment 1 in which infants looked longer at the face versus the ball across conditions, infants would not show a preference for the inverted doll in Experiment 2. We averaged infants’ log looking times across test trials in which the ball was revealed and test trials in which the inverted doll was revealed, then conducted a repeated-measures ANOVA with Object Revealed (ball or inverted doll) as a within-subjects factor and Condition (control or swap) as a between-subjects factor. We found no main effect of Object Revealed (F(1,22) Conceptually rich, perceptually sparse object representations 17 < 1, p = .98, ηp2 < .001) suggesting that infants had no overall preference for the inverted doll (raw Mcontrol = 5.66 s; raw Mswap = 4.14 s) over the ball (raw Mcontrol = 4.69 s; raw Mswap = 4.80 s). Bayes factor analysis yielded moderate support for the null hypothesis (JZS BF01 = 4.66). Critically, we also observed no Object Revealed X Condition interaction (F(1,22) = 1.74, p = .20, ηp2 .073), suggesting that infants’ pattern of looking did not vary across conditions. Finally, we compared infants’ log looking times to control and swap outcomes across Experiments 1 and 2, using a repeated-measures ANOVA with Trial (1st, 2nd, 3rd, or 4th) as a within-subjects factor and Condition (control or swap) and Experiment (1 or 2) as between-subjects factors. We found no main effect of Trial (F(3,132) < 1, p = .94, ηp2 = .003), and no Trial X Condition interaction (F(3,132) < 1, p = .84, ηp2 = .006) or Trial X Experiment interaction (F(3,132) < 1, p = .46, ηp2 = .02). There was also no main effect of Condition (F(1,44) < 1, p = .45, ηp2 = .013) and no main effect of Experiment (F(1,44) = 3.63, p = .06, ηp2 = .076). Of critical interest is the Condition X Experiment interaction, which over all 4 test trials was weak (F(1,44) = 3.44, p = .07, ηp2 = .072). However, inspection of trial-by-trial plots (see Figures S1 and S2) shows that, whereas all trials, with the exception of Trial 4, displayed evidence of an interaction, effects were largest on first test trial looking times and diminished on later trials, as is typical in infant looking time studies with multiple test trials. An ANOVA conducted on first trial log looking times showed no main effect of Condition (F<1) but did show a main effect of Experiment (F(1,44) = 4.95, p = .031, ηp2 = .101) subsumed under a significant Condition x Experiment interaction (F(1,44) = 4.74, p = .035, ηp2 = .097). No other significant effects were found on later trials. Conceptually rich, perceptually sparse object representations 18 We followed these analyses with planned comparisons (Hays, 1998; Chapter 11), using independent samples t-tests and Bayes factor analysis on infants’ log looking times averaged across all test trials. We found that infants looked significantly longer in the swap condition in Experiment 1 versus Experiment 2 (t(22) = 2.84, p = .01, 95% CI [.053, .338]; BF10 = 5.339), and looked roughly equally in the control conditions (t(22) = .033, p = .97, 95% CI [-.186, .194]; BF01 = 2.68). We then collapsed the two control conditions and compared the resulting three groups (Human Swap vs Inverted Swap vs Controls) using Dunnett’s t-test for multiple comparison against a single control with predicted direction, finding a significant difference between Human Swap vs. Controls (p = .032) but not Inverted Swap vs Controls (p = .89). Discussion Unlike infants in Experiment 1, infants in Experiment 2 did not look longer at the swap outcome, suggesting that they failed to notice when the first-hidden object changed. The only systematic difference between Experiments 1 and 2 was that the doll was inverted. Together, these results suggest that when objects contrast in categorical identity, infants may successfully encode these identities. However, Bayes Factor analyses suggested only weak to moderate effects. To confirm the observed pattern of results, in Experiment 3, we conducted a critical replication of the two primary conditions of interest from Experiments 1 and 2: upright swap and inverted swap, using a larger sample of infants. Experiment 3: Replication: Upright swap versus inverted swap Participants Conceptually rich, perceptually sparse object representations 19 A separate group of 30 healthy, full-term infants participated (mean age: 5 months 28 days; range: 5 months 0 days to 7 months 4 days; 16 girls). Infants were randomly assigned to either the upright swap condition (n = 15) or the inverted swap condition (n = 15; between subjects, design). An additional two infants were tested but excluded from analysis due to experimenter error (1) or parental interference (1). Materials and Procedure Stimuli and procedure were identical to the swap trials of Experiments 1 and 2. Infants in the upright swap condition observed familiarization and test trials consistent with the swap outcome of Experiment 1 (Figure 1). Infants in the inverted swap condition observed familiarization and test trials consistent with the swap outcome of Experiment 2 (Figure 4). A naïve observer, who was unable to see what infants were seeing, measured looking time during test trials. Infants’ looking times were later recoded offline by two separate observers who were naïve to condition. Inter-observer agreement was high (mean r = .93). Results Data were first log-transformed to correct for right skew. A repeated-measures ANOVA with Trial (1st, 2nd, 3rd, or 4th) as a within-subjects factor and Condition (upright swap or inverted swap) as a between-subjects factor revealed no main effect of Trial (F(3,84) = 1.31, p = .28, ηp2 = .05) and no Trial X Condition interaction (F(3,84) = 1.04, p = .38, ηp2 = .04). We observed a significant main effect of Condition (F(1,28) = 18.926, p < .001, ηp2 = .40); infants looked significantly longer at the upright swap outcome than the inverted swap outcome (Figures 2 and 3). Bayes factor analysis yielded decisive Conceptually rich, perceptually sparse object representations 20 evidence for the alternative (JZS BF10 = 138.4). Analysis of the first test trial only revealed a similar pattern (t(28) = 2.51, p = .02; JZS BF10 = 3.27; Figure S3). Next, we asked whether the significant main effect of condition could be driven by a preference to look at the upright face in the upright swap condition, rather than a detection of a change to the object’s identity. We computed infants’ mean log looking times on test trials in which the ball was hidden and the doll was revealed and on test trials in which the doll was hidden and the ball was revealed. We submitted these to a repeated measures ANOVA with Object Revealed (ball or doll) as a within subjects factor and Condition (upright swap or inverted swap) as a between subjects factor. We again observed a main effect of Condition (F(1,28) = 18.92, p < .001, ηp2 = .403), but no main effect of Object Revealed (F(1,28) < 1, p = .6, ηp2 = .01): infants looked very slightly, but not significantly, longer at the doll in the upright swap condition (raw Mdoll = 12.77 s; raw Mball = 9.51 s; t(14) = -2.09, p = .06, JZS BF10 = 1.29), and looked roughly equally at the two outcomes in the inverted swap condition (raw Mdoll = 5.81 s; raw Mball = 6.62 s; t14 = .80, p = .44, JZS BF01 = 3.85). We also observed no Object Revealed X Condition interaction (F(1,28) = 3.39, p = .08, ηp2 = .11), suggesting that the main effect of condition was not driven by mere preference for upright faces in the upright swap condition. We then compared infants’ log looking times to each outcome in Experiment 3 to infants’ log looking times to the swap outcomes of Experiments 1 and 2. We used a repeated measures ANOVA with Trial (1st, 2nd, 3rd, or 4th) as a within subjects factor and Condition (upright swap or inverted swap) and Experiment Type (original or replication) as between subjects factors. We again found no main effect of Trial (F(3,150) < 1, p Conceptually rich, perceptually sparse object representations 21 =.733, ηp2 =.008) and no Trial X Condition interaction (F(3,150) = 1.10, p =.35, ηp2 =.022). There was also no Trial X Experiment Type interaction (F(3,150) < 1, p =.502, ηp2 =.016) and no three-way interaction (F(3,150) = 1.378, p =.252, ηp2 =.027). Critically, we observed a significant main effect of Condition (F(1, 50) = 27.22, p <.001, ηp2 =.353); infants looked significantly longer at the upright swap outcome than the inverted swap outcome, with Bayes factor analysis yielding decisive support for the alternative (JZS BF10 on combined mean log looking times = 721.59; JZS BF10 on combined log looking times on trial 1 only = 45.5; Figure 3). There was a significant main effect of Experiment Type (F(1, 50) = 15.48, p <.001, ηp2 =.236) but no Condition X Experiment Type interaction (F(1, 50) < 1, p =.50 ηp2 =.009); overall, infants looked longer in Experiment 3 (replication) versus the original experimental conditions, but their pattern of looking longer at the upright swap versus the inverted swap outcome did not differ across experiments. Finally, to confirm the results of Experiments 1 and 2, we conducted separate Bayes Factor analyses on the combined upright swap conditions from Experiments 1 and 3 versus the combined control conditions from Experiments 1 and 2, and the combined inverted swap conditions from Experiments 2 and 3 versus the combined controls. Analysis of infants’ mean log looking times in the combined upright swap conditions versus the control conditions yielded JZS BF10 = 2288.3, with a similar pattern observed on the first trial only (JZS BF10 = 1632.1), decisively favoring the alternative hypothesis. Analysis of infants’ mean log looking times in the combined inverted swap conditions versus the controls yielded JZS BF01 = 3.2, with a similar pattern observed on the first trial only (JZS BF01 = 3.4), suggesting moderate evidence in favor of the null. Conceptually rich, perceptually sparse object representations 22 Discussion The results of Experiment 3 replicated the pattern of results observed in Experiments 1 and 2. Together, these results suggest that when objects contrast in categorical identity, infants are able to track their identities into occlusion. One additional possibility should be explored: presenting infants with objects from two distinct conceptual categories may help infants to better encode the features of each object in addition to, or even instead of, the objects’ categorical identities. Infants in Experiments 1 and 3 could have tracked the objects by recalling the features of the objects, the categories of the objects, or both. To test this possibility, in Experiment 4 we again hid two categorically distinct objects as in Experiment 1. However, on swap trials, the first-hidden object was revealed to have changed to a featurally distinct object from the same category as the hidden object. If the categorical distinctions between the objects prompted infants to encode the features of the objects, infants should successfully detect the large featural change between objects with the same categorical identity. However, if the categorical distinctions between the objects prompted infants to encode the categorical identity but not the featural identity of the probed object, infants should fail. Experiment 4: Human versus Object, Within-Category Change Participants A separate group of 24 healthy, full-term infants participated (mean age: 6 months 7 days; range: 5 months 5 days to 7 months 13 days; 12 girls). Infants were randomly assigned to either the control condition (n = 12) or the swap condition (n = 12; between Conceptually rich, perceptually sparse object representations 23 subjects design). An additional four infants were tested but excluded from analysis due to fussiness (3) or experimenter error (1). Materials Materials were identical to Experiment 1, except two additional objects were used: a green ball with red polka dots (~9 cm in diameter) glued onto a 4.5 x 4.5 black foam-core platform, and a doll’s head with plastic blonde hair, pink skin, blue eyes, and a pink necklace, also glued onto a 4.5 x 4.5 black foam-core platform, and roughly the same size as the other objects used. These objects were chosen to be highly featurally dissimilar from their counterparts from the same category (see Figure 5). Procedure Familiarization Trials. The four familiarization trials proceeded similarly to Experiments 1 and 2. On each familiarization trial, infants saw one ball and one doll placed on the stage. Across trials, infants saw each object from the one category (humanlike dolls) paired with each object from the other category (nonhuman object balls), with order and side of placement counterbalanced across trials and across infants. Thus, while infants got less overall exposure to each object than in Experiments 1 and 2, they experienced each object in each potential location, in each potential presentation order, and paired with each object from the other category during familiarization trials. Figure 5 (left panel) shows examples of three of the four familiarization trials. See also Supplementary Video S7 for an example familiarization trial. Test Trials. The four test trials proceeded similarly to Experiments 1-3. A ball and a doll were placed one at a time on the stage. After 4 s, the experimenter hid these sequentially behind the occluders. The experimenter then lifted the screen occluding the Conceptually rich, perceptually sparse object representations 24 first-hidden object. Half of infants saw the object that was hidden there originally (e.g. striped ball hidden, striped ball revealed, control condition). The other half of infants saw that the object had swapped for a featurally distinct object of the same category, (e.g. striped ball hidden, green ball revealed, swap outcome). Infants then saw three more test trials, for a total of four test trials. Across test trials, as in familiarization trials, infants always saw one ball and one doll on each trial, with the order and side of placement of the objects counterbalanced across trials and infants. Thus, on half the trials, the ball was hidden first, and on the other half the doll was hidden first. Figure 5 (right panel) shows a sample test trial for Experiment 4. See also Supplementary Videos S8 and S9 for example control and swap outcome test trials, respectively. Conceptually rich, perceptually sparse object representations 25 Figure 5. Left panel shows examples of three of the four familiarization trials for Experiment 4. Infants observed all possible pairings of dolls and balls, with order and side of placement counterbalanced across trials and across infants. Right panel shows an example of one of the four test trials for Experiment 4. Infants observed a doll and a ball hidden alternately behind separate occluders. The occluder hiding the first-hidden object was then removed to reveal either the object that had been hidden in that location originally (control outcome) or the object from the same category to which infants had been familiarized, but which was not present on that trial (e.g. yellow ball hidden, green ball revealed; swap outcome). A naïve observer who was unable to see what infants were seeing measured looking time during test trials. Infants’ looking times were later recoded offline by an observer who was naïve to condition, and these looking times were analyzed. A second Conceptually rich, perceptually sparse object representations 26 naïve observer also coded a random 50% of infants. Inter-observer agreement was high (r=.98). Results As in Experiments 1-3, data were log-transformed to correct for right skew. A repeated-measures ANOVA with Trial (1st, 2nd, 3rd, or 4th) as a within-subjects factor and Condition (control or swap) as a between-subjects factor revealed no main effect of Trial (F(3,66) = 1.88, p = .14, ηp2 = .08), no Trial X Condition interaction (F(3,66) < 1, p = .76, ηp2 =.02). There was also no main effect of Condition (F(3,66) < 1, p = .95, ηp2 < .001), suggesting that infants looked equally at control and swap outcomes, with Bayes factor analysis yielding weak evidence for the null (JZS BF01 = 2.19; Figure 2). Analysis of the first trial only revealed a similar pattern (t(22) = -.418, p = .680; JZs BF01 = 2.513; Figure S4). We also asked whether infants’ pattern of looking to the different objects (e.g. a preference for the doll) impacted their looking across conditions. A repeated-measures ANOVA with Object Revealed (ball or doll) as a within-subjects factor and Condition (swap or control) as a between-subjects factor revealed that infants tended to look longer at dolls (Mcontrol = 9.36 s; Mswap = 8.58 s) than at balls (Mcontrol = 7.78 s; Mswap = 5.39 s) but this difference was not significant (F(1,22) = 3.43, p = .08, ηp2 =.14). There was also no interaction between Object Revealed and Condition (F(1,22) = 1.33, p = .26, ηp2 = .06) suggesting that infants’ pattern of looking to the two object types did not differ as a function of condition. Conceptually rich, perceptually sparse object representations 27 Finally, we asked whether infants’ looking time to the within-category swap outcome of Experiment 4 differed significantly from their pattern of looking to the swap outcomes in Experiments 1-3. This allowed us to test the critical hypothesis that infants would be able to detect a change to the identity of the first-hidden object only if the following two criteria were met: a) the hidden objects contrasted in category and b) the first-hidden object was revealed to have changed to an object from a different category. If this hypothesis is supported, it would suggest that infants encode the category of the firsthidden object but not necessarily its specific featural identity. We hypothesized that infants would look significantly longer at the upright swap outcomes of Experiments 1 and 3 versus the inverted swap outcomes of Experiments 2 and 3 or the within-category swap outcome of Experiment 4. We submitted infants’ log looking times to the swap outcomes of Experiments 1-4 to a repeated measures ANOVA with Trial (1st, 2nd, 3rd, or 4th) as a within subjects factor and Condition (upright swap, inverted swap, or within-category swap) and Experiment Type (original or replication) as between subjects factors. Main effects were examined using planned post hoc tests with Bonferroni correction for multiple pairwise comparisons. We observed no main effect of Trial, and no interactions between Trial and any other variables (all F < 1.53, all p > .05). There was a significant main effect of Experiment Type (F(1,61) = 16.10, p < .001, ηp2 =.209), but no Experiment X Condition interaction (F(1,61) < 1, p = .49, ηp2 =.008); infants looked longer overall in the replication versus the original experiments (p < .001), regardless of outcome. Critically, we observed a significant main effect of Condition (F(1,61) = 15.12, p <.001, ηp2 =.331). Post hoc pairwise comparisons with Bonferroni correction revealed Conceptually rich, perceptually sparse object representations 28 that infants looked significantly longer at the upright swap outcome versus both the inverted swap outcome (p < .001) and the within category swap outcome (p = .05), with Bayes factor analysis yielding decisive (upright vs. inverted swap JZS BF10 = 721.59) and moderate (upright vs. within-category swap JZS BF10 = 3.80) support for the alternative hypothesis. Infants’ looking times to the inverted swap and within category swap outcomes did not differ significantly (p =.237), with Bayes factor analysis offering anecdotal support for the null (JZS BF10 = 1.44). General Discussion We tested the hypothesis that the conceptual content of object representations may be richer than the perceptual content. We tested 6-month-old infants, who can individuate multiple objects but can remember the featural identity (e.g. shape) of only a single object (Káldy & Leslie, 2005; Kibbe & Leslie, 2011, 2016). We hid two objects from infants and tested their ability to recall the first-hidden of the two objects, for which infants typically fail to recall perceptual features. In Experiment 1, we hid two objects from different conceptual categories (human and non-human object) and found that 6month-olds’ pattern of looking suggested that they successfully recalled the categorical identity of the probed object. In Experiment 2, we ruled out lower-level perceptual explanations for infants’ pattern of looking in Experiment 1. In Experiment 3, we replicated the critical comparison from Experiments 1 and 2 with a larger sample of infants. Finally, in Experiment 4, we found that presenting infants with categorically distinct objects did not simply improve their working memory for the featural identities of those objects. Conceptually rich, perceptually sparse object representations 29 Together, our results suggest that categorical identity may be encoded more readily than featural identity in the object representation. This suggests that (some) conceptual content may be a privileged or essential part of the structure of object representations, as hypothesized by Baillargeon et al. (2012). Our results are consistent with previous work on object individuation, showing that, by 10 months, if infants are sensitive to conceptual distinctions between the objects they will use these distinctions to form representations of distinct individuals (e.g. Bonatti et al., 2002; Futo et al., 2010; Needham et al., 2005; Xu & Carey, 1996). However, in those studies infants did not need to encode or retain categorical identity in their representations of the objects. Our results show for the first time that such information finds its way into the object representation itself, and that this happens already by 6 months. Yoon, Johnson and Csibra (2008) argued that a social communication context, such as ostensive pointing, is important for 9-month-olds to remember object type but at the cost of remembering location. Our findings with 6-month-olds show both memory for location and for object type even outside of a communicative context. Core concept binding? What is the structure of infants’ representations of occluded objects? Perhaps some conceptual information is essentially represented whenever an object is (e.g. Baillargeon et al., 2012). For 6-month-olds, whose conceptual repertoire is small, essential content may be limited to “core knowledge” concepts such as OBJECT and AGENT (c.f. Carey, 2009; Leslie, 1994; Spelke et al., 1992). The conceptual content included automatically in representations of occluded objects may be limited to these Conceptually rich, perceptually sparse object representations 30 core-knowledge concepts across the lifespan, or may be expanded to include conceptual content acquired through specific experience, including kind-based (Feigenson & Halberda, 2008; Needham et al., 2005; Xu & Carey, 1996) and functional (e.g. Futo et al., 2010; Stavans & Baillargeon, 2016) information. A different possibility is that conceptual information requires binding to object representations, just like perceptual features do. However, concept bindings may be easier to form, cheaper to maintain, or more durable over time, and certain concepts, like HUMAN, may be privileged in certain ways (Bonatti et al., 2002; Stein, Sterzer, Peelen, 2012). For example, while feature binding may rely on sensory traces in early visual areas (e.g. Harrison & Tong, 2009; Serences et al., 2009) — traces that are inherently noisy and subject to interference as new sensory information comes in — concept binding may rely on more durable long-term representations that are more resilient to interference or decay. In either case, the notion of concept binding is consonant with previous work suggesting that an object representation can be a raw, featureless pointer to an occluded object’s location in space, containing no information about the object’s identity (conceptual or otherwise) (Kibbe & Leslie, 2011; Leslie, et al., 1998; Pylyshyn, 1989; Scholl & Leslie, 1999).