Linguistic Labels, Dynamic Visual Features, and Attention in Infant Category Learning

Wei (Sophia) Deng; Vladimir M Sloutsky

doi:10.1016/j.jecp.2015.01.012

. Author manuscript; available in PMC: 2016 Jun 1.

Published in final edited form as: J Exp Child Psychol. 2015 Mar 25;134:62–77. doi: 10.1016/j.jecp.2015.01.012

Linguistic Labels, Dynamic Visual Features, and Attention in Infant Category Learning

Wei (Sophia) Deng ¹, Vladimir M Sloutsky ¹

PMCID: PMC4394365 NIHMSID: NIHMS675698 PMID: 25819100

Abstract

How do words affect categorization? According to some accounts, even early in development, words are category markers and are different from other features. According to other accounts, early in development, words are part of the input and are akin to other features. The current study addressed this issue by examining the role of words and dynamic visual features in category learning in 8- to 12- month infants. Infants were familiarized with exemplars from one category in a label-defined or motion-defined condition and then tested with prototypes from the studied category and from a novel contrast category. Eye tracking results indicated that infants exhibited better category learning in the motion-defined than in the label-defined condition and their attention was more distributed among different features when there was a dynamic visual feature compared to the label-defined condition. These results provide little evidence for the idea that linguistic labels are category markers that facilitate category learning.

The ability to form categories is an important component of human cognition that appears early in development: infants exhibit evidence of category learning during the first months of life (Quinn et al., 1993; Younger & Cohen, 1985). There is also evidence suggesting that language may affect this process, although the mechanisms underlying the effects of language remain a matter of debate.

Some suggested that words accompanying category members have the special status of category markers and, as such, they guide or supervise category learning in infancy (Waxman & Markow, 1995; see also Westermann & Mareschal, 2014). At the same time, others suggested that early in development words are akin to other features, but they may become category markers in the course of development (Gliozzi et al., 2009; Sloutsky, 2010; Sloutsky & Lo, 1999; Sloutsky & Fisher, 2004; Sloutsky et al., 2001). As we discuss below, distinguishing between these positions has profound consequences for our understanding of the relationships between language and cognition and the nature of learning early in development.

According to the former theory, “infants embark on the task of word learning equipped with a broad, universally shared expectation, linking words to commonalities among objects” (Waxman, 2003, p. 220). As a result, words, but not other kinds of auditory input, facilitate infants’ category learning by attracting attention to within-category commonalities (Waxman & Markow, 1995; Waxman & Booth, 2001), thus effectively supervising category learning. These effects are supervisory because labels guide learning by attracting attention to commonalities.

There is some evidence consistent with this view. First, words may facilitate infants’ categorization above and beyond other kinds of auditory input (Balaban & Waxman, 1997; Fulkerson & Haaf, 2003; Ferry et al., 2010). Second, facilitative effects of words were reported for basic-level as well as at superordinate or global levels (Balaban & Waxman, 1997; Waxman & Booth, 2003; Waxman & Markow, 1995). Third, there are reports that facilitative effects of labels are specific rather than general in nature: count nouns and adjectives have initially similar effects on category learning, whereas around 14-months of age count nouns are more likely to facilitate category learning than adjectives (Waxman & Booth, 2001). This finding suggests that count nouns may play a special role in category learning. And finally, labels may facilitate property induction above other kinds of input (Keates & Graham, 2008).

There are challenges, however, to the idea that words are category markers in infancy. First, even if words affect category learning in infancy, they do not have to function as category markers supervising learning, but can be instead part of the stimulus input and influence learning in a bottom-up fashion. For example, Plunkett et al (2008) presented 10-month-old infants with a category-learning task, such that the to-be-learned category consisted of two clusters of artificial creatures (i.e., a broad category somewhat analogous to a global category encompassing cats and horses). When the category was presented in silence, participants learned two narrow categories, whereas when one common label accompanied each item, participants learned the single broad category. Although it is tempting to conclude that these results indicate that labels supervised category learning, this conclusion is unwarranted. Specifically, when Gliozzi et al (2009) modeled data reported by Plunkett et al (2008) using self-organizing maps, a model that assumed that labels are features and function as input rather than top-down supervisory signals was able to account for the reported pattern.

Second, findings that labels facilitate infant category learning are tenuous at best – facilitation transpires in some studies and does not transpire in others. This is because many studies compared the effects of labels with those of unfamiliar sounds, but not with a silent condition. When a silent baseline was introduced (e.g., Robinson & Sloutsky, 2007), labels were not found to facilitate infants’ category learning above the silent baseline.

And finally, even studies demonstrating facilitative effects of labels have generated inconsistent findings regarding the age at which labels facilitate category learning. For example, Booth and Waxman (2002) demonstrated that for an artifact category, words alone facilitated category learning only at 18-months of age, whereas around 14-months of age, words had to be paired with object function to facilitate learning above the baseline. It was argued that both words and functions (a) indicate human agency and (b) highlight commonalities among objects. Results of this study seem to be in a sharp contrast with studies where words were claimed to facilitate category learning in very young infants. In particular, Ferry et al (2010) found evidence that labels facilitate category learning in 3- to 4-month-olds. Why do labels facilitate category learning in 3–4 months-olds, while failing to facilitate learning in much older infants (e.g., Booth & Waxman, 2002; Robinson & Sloutsky, 2007, 2008)?

According to the label-as-feature proposal, at least early in development, words are part of input – they are features of items rather than category markers (Gliozzi et al., 2009; Sloutsky & Lo, 1999; Sloutsky & Fisher, 2004; Sloutsky et al., 2001). For example, there is evidence that early in development auditory input overshadows (or attenuates processing of) corresponding visual input (Lewkowicz, 1988a, 1988b; Napolitano & Sloutsky, 2004; Robinson & Sloutsky, 2004a, 2004b, 2008; Sloutsky & Napolitano, 2003). Therefore, under most conditions words do not facilitate infant category learning, but, under some conditions, they may interfere with learning.

Of course, even if words start out as features, they may become category markers later in development (e.g., Deng & Sloutsky, 2012). For example, Deng and Sloutsky (2012) used a variant of Yamauchi and Markman’s (1998, 2000) category learning paradigm to teach preschoolers and adults two novel categories. During training, the category label correlated perfectly with a pattern of motion, which is a highly salient visual feature (Egeth & Yantis, 1997, for review), whereas the rest of the category features were probabilistic. At test, the label was pitted against the pattern of motion and participants had to predict one of the features of a test item. The researchers found that unlike many adults, who relied on a category label, children relied on the salient feature, despite the fact that their memory for the label was as good as that of adults. These results raised interesting questions. If a visual feature that is a part of input has a greater effect on category learning than a label, what makes the label a category marker? And if words are not category markers for preschoolers, how can they be category markers for infants?

In sum, there are two theoretical positions with respect to the role of words in category learning, which differ with respect to the underlying mechanism and a developmental trajectory. Distinguishing between these possibilities and understanding the mechanisms underlying the effect of words on category learning is of critical importance for understanding cognitive development. If from early in development words function as category markers supervising learning, then top-down effects may play a significant role in early cognitive development. Also, given that supervision results in the ability to learn substantially more complex categories than unsupervised learning (Rumelhart, 1989), if words are supervisory signals for infants, our construal of what infants can and cannot learn will be subject to substantial revision.

Effects of Words and Other Features on Category Learning in Infancy: Potential Mechanisms

Although the reviewed evidence is controversial with respect to the idea that labels are category markers, none of the reviewed studies examined the mechanisms hypothesized by each theoretical position. At the same time, each position gives divergent predictions as to how words should affect category learning in comparison with other features. According to the label-as-category-marker position, attracting attention to commonalities is the mechanism via which labels facilitate category learning (e.g., Waxman, 2003). Therefore, when labels are presented, category learning should be accompanied by some form of attention optimization (Blair et al. 2009; Hoffman & Rehder, 2010) — diffused attention early in learning and followed by shifting attention to within-category commonalities. Furthermore, attention optimization when labels are present should be greater than in no-label conditions, and this attention optimization should lead to better category learning.

In contrast, according to the label-as-feature position, (a) labels should not attract attention to commonalities and (b) due to possible auditory overshadowing effects, labels should either interfere with infant category learning or not affect it at all. Furthermore, there is evidence that successful category learning in 6–8-months-old infants is accompanied not by increased attention to commonalities (which is often the case in adults), but by broader exploration of features and thus more broadly distributed attention (e.g., Best, Yim, & Sloutsky, 2013). This prediction is based on a substantial body of literature demonstrating that distributed attention (with shorter individual fixations and frequent gaze shifts) is generally associated with better learning in infancy (e.g., Bronson, 1991; Colombo et al., 1991; Jankowski et al., 2001; Rose et al., 2003). There is also evidence that distributed attention is not merely associated, but it leads to better learning. In one study (Jankowski et al., 2001), researchers introduced a peripheral dynamic cue (a dynamic engager appearing in various parts of the screen) to induce shorter fixations, multiple gaze shifts, and thus a more distributed pattern of attention in 5-month-old infants. Results indicated that training resulted in more distributed attention (compared to no training) and in more efficient learning of presented items. This finding suggests that dynamic visual features (especially those that appear peripherally) may encourage distributed attention, thus facilitating learning. Given that successful category learning in infancy is also accompanied by distributed attention (e.g., Best et al., 2013), it is possible that such dynamic visual features may also affect infant category learning.

Why do dynamic visual features encourage distributed attention? In contrast to static visual features that maintain their salience throughout the trial, dynamic visual features become highly salient only during the period when they are dynamic. As a result, if a dynamic visual feature is presented peripherally, at the very minimum participants would need to move their gaze from the point of their initial fixation to the moving part and then, when the motion ceases to the point of their final fixation. Each of these gaze shifts may be accompanied by short fixations (and thus attention to) multiple regions of the stimulus (see Jankowski et al., 2001, for a discussion).

Therefore, compared to labels, these dynamic visual features may result in more gaze shifts, more distributed attention, and better, more robust learning. This prediction contrasts sharply with that of the label-as-category-marker position, which predicts that labels facilitate category learning by attracting attention to commonalities. The goal of the current research was to test this prediction by using a combination of eye tracking and a more traditional novelty preference paradigm.

Overview of Current Study

The study reported here presents two experiments designed to examine the effects of labels and dynamic visual features on category learning in infancy. Experiment 1 included two between-subjects conditions: label-defined condition and motion-defined condition. In both conditions, infants were familiarized with exemplars from one category and then tested with the prototype of this category and that of the contrast category. Infants saw the same testing stimuli in both conditions, with neither label nor motion being presented during testing. Experiment 2 was aimed at further comparing the effects of labels with a silent control condition. In both experiments, eye gaze data were collected. Recall, that, according to the label-as-category-marker position, labels should (a) facilitate category learning compared to the other conditions and (b) lead to greater attention to commonalities than in the other two conditions. In contrast, according to the label-as-feature view, (a) because of auditory overshadowing, labels should not facilitate category learning and (b) more robust category learning may be accompanied by more distributed attention

EXPERIMENT 1

Method

Participants

Fifty-one infants (30 girls) ranging in age from 8 to 12 months (M = 10 months, 11 days; SD = 1 month, 20 days) were recruited. Ten infants were excluded from the analyses (two due to fussiness and eight for not looking at a single test trial).

Apparatus

A Tobii T60 eye-tracker with the sampling rate of 60 Hz was used to collect eye gaze data. The eye-tracker was integrated into a 17-inch computer monitor and located on a table inside a booth enclosed by black curtains. A trained experimenter monitored the experiment using Tobii Studio gaze analysis software installed on a 19-inch Dell OptiPlex 755 computer outside the booth. A video stream displaying participants’ activities was projected onto a 9-inch black and white Sony SSM-930 CE television for the experimenter’s online monitoring. Two Dell speakers were located behind a black curtain on each side of the eye-tracker.

Materials and Design

The materials were colorful drawings of artificial creatures and novel labels “flurp” and “jalet” (see Figure 1A). The creatures had five features varying in color and shape and consisted of two categories. As shown in Table 1, the categories had a family-resemblance structure, which was derived from two prototypes (A0 and B0) by modifying the values of one of five features – head, antennae, hands, body, or feet. We used these novel categories to ensure that none of the infants was familiar with these categories prior to the experiment and all participants had to learn these categories de novo. Another advantage of these stimuli is that they have been extensively tested in previous work with preschoolers (Deng & Sloutsky, 2012; 2013).

Example Stimuli. A. Prototypes of stimuli from Categories A and B; B. Procedures used in this study. C. Areas of interest (AOIs) shown as gray ellipses.

Table 1.

Category structure used in training.

Category A							Category B
Stimuli	Head	Body	Hands	Feet	Antenna	Label/Motion	Stimuli	Head	Body	Hands	Feet	Antenna	Label/Motion
A1	1	1	1	1	0	1	B1	0	0	0	0	1	0
A2	1	1	1	0	1	1	B2	0	0	0	1	0	0
A3	1	1	0	1	1	1	B3	0	0	1	0	0	0
A4	1	0	1	1	1	1	B4	0	1	0	0	0	0
A5	0	1	1	1	1	1	B5	1	0	0	0	0	0
A0	1	1	1	1	1	1	B0	0	0	0	0	0	0

Open in a new tab

Note. The value 1 = any of five dimensions identical to “Flurp” (see Figure 1). The value 0 = any of five dimensions identical to “Jalet” (see Figure 1). A0 and B0 are prototypes of each category.

There were two between-subjects conditions: label-defined and motion-defined. In the label-defined condition, the label presented during training was the same for all the exemplars, whereas in the motion-defined condition the pattern of motion presented during training was the same. To create a dynamic visual feature, the feet were animated using Macromedia Flash MX software. For all “flurps” the feet stretched up and down, whereas for all “jalets” the feet moved sideways. Because the goal was to examine effects of labels and dynamic visual features on category learning, neither labels nor patterns of motion were presented at test.

Procedure

Infants were seated on parents’ laps approximately 60 cm away from the eye-tracker. Parents were instructed not to interact with infants and to avoid speaking or pointing. Prior to the experiment, infants completed a 5-point calibration sequence. The calibration consisted of dynamic kitten images accompanied by a “bouncing” sound appearing in different locations on the screen.

The experiment proper consisted of 20 familiarization and 4 test trials. The trials were mixed and pseudo-randomly assigned to four blocks, with each block consisting of five familiarization trials followed by one test trial (see Figure 1B for the overall sequence in a block). On each familiarization trial, infants saw a creature generated from the same category (one of the categories shown in Table 1) on a white background lasting for 8000 ms and heard a phrase starting at the onset of each trial. A subset of infants studied Category A whereas the rest of the infants studied Category B. In the label-defined condition, the phrase (e.g., “Look! This is a Flurp”) was presented at the beginning of each trial and lasted for approximately 2800 ms, whereas, in the motion-defined condition, the phrase did not include the label (e.g., “Look at this one!”). The feet of the creature started moving after the phrase ended and the motion lasted for 3000 ms. The onset of motion was approximately the same as that of the label in the label-defined condition (i.e., 2300 ms into the trial).

Each test trial lasted for 8000 ms and presented a pair of items – the prototype of the studied category and the prototype of the non-studied category, with the left-right position of each prototype counterbalanced across test trials. Note that neither prototype was presented during training. These test items (these were the same across the conditions) were different from training trials as they were presented without either label or motion. A dynamic bouncing ball was presented between trials within each block, whereas a short and task-irrelevant cartoon video was presented between blocks to maintain engagement of infant participants. All gaze data were recorded by the computer using Tobii Studio gaze analysis software.

Results and Discussion

Gaze data were exported using Tobii Studio gaze analysis software. For each stimulus seven areas of interest (AOIs) for fixations were defined: ellipses (ranging in visual angle from 2.4° by 2.4° to 3.8° by 5.7°) surrounding each feature of the creatures. Given that hands, feet, and antennae appeared in pairs, each pair was treated as a single AOI, which resulted in five AOIs used for data analyses: head, body, hands, feet, and antennae (see Figure 1C for AOIs). The gaze durations were weighted by the area of each AOI. The analyses focused on (1) looking time at familiarization; (2) novelty preference score based on the proportion of looking time to the prototype of the novel category as compared to the total looking time to both of the prototypes at test; and (3) patterns of attention during test and familiarization based on (a) the proportion of looking time to different features and (b) the number of gaze shifts between AOIs.

Looking Time at Familiarization

To ascertain that the levels of engagement with the stimuli were comparable across the conditions, we compared accumulated looking time to the stimuli on familiarization trials in the label-defined condition with that in the motion-defined condition. The accumulated looking time data were submitted to a 4 (Block: 1 vs. 2 vs. 3 vs. 4) by 2 (Condition: label-defined vs. motion-defined) mixed ANOVA, with block as a within-subjects factor and condition as a between-subjects factor. Results revealed a main effect of block, F(3, 117) = 37.26, MSE = 0.02, p < .01, η_p² = 0.296, with infants’ accumulated looking time decreasing through blocks. However, infants’ accumulated looking time did not differ between the label-defined and motion-defined conditions, p = .52. These results indicated that infants exhibited comparable levels of engagement across the two conditions.

To further examine whether infants became more familiarized with stimuli in the course of learning, we compared the average looking time on the first-three and the last-three familiarization trials in the label-defined condition with that in the motion-defined condition (see Figure 2). These data were submitted to a 2 (Trial: first-three vs. last-three) by 2 (Condition: label-defined vs. motion-defined) mixed ANOVA, with trial as a within-subjects factor and condition as a between-subjects factor. There was a main effect of trial, F(1, 39) = 57.78, p < .001, η_p² = 0.597, with infants in both conditions decreasing looking during familiarization, paired-samples ps < .001. There was a significant trial by condition interaction, F(1, 39) = 4.62, p = .038, η_p² = 0.106. The interaction indicated that infants in the label-defined condition started with longer looking on the first-three trials than those in the motion-defined condition, independent-samples t(31.7) = 2.44, p = .021, d = 0.75, whereas there was no difference in the last three trials, p = .629. Therefore, infants in the label-defined condition exhibited somewhat greater familiarization than infants in the motion-defined condition.

Average looking time on the first-three and last-three familiarization trials in the label-defined and motion-defined conditions.

Novelty Preference

To examine how labels or patterns of motion affected infants’ categorization, a novelty preference score was calculated for each test trial. Because there was only one test trial per block, we averaged test trials for blocks 1–2 and for blocks 3–4 (see Figure 3). The data in Figure 3 were submitted to a 2 (Block: 1–2 vs. 3–4) by 2 (Condition: label-defined vs. motion-defined) mixed ANOVA, with block as a within-subjects factor and condition as a between-subjects factor. There was a significant main effect of condition, with infants having higher novelty preference scores in the motion-defined condition, F(1, 39) = 5.78, MSE = 0.12, p = .021, η_p² = 0.129. Neither the effect of block (p = .079), nor the interaction (p = .325) reached significance.

In addition, infants in the motion-defined condition exhibited above-chance novelty preference in Blocks 3–4, one-sample t(19) = 2.87, p = .010, d = 0.64, but not in Blocks 1–2, p = .949; whereas in the label-defined condition, novelty preference was not different from chance in either block, both ps > .251. Therefore, after four blocks of training participants in the motion-defined condition exhibited evidence of category learning, whereas participants in the label-defined condition failed to learn.

Distributions of individual novelty preference scores by condition are presented in Figure 4. As shown in Figure 4, in the first half of training (Blocks 1–2) there were comparable numbers of participants exhibiting novelty preference in either condition, whereas in the second half of training (Blocks 3–4), there were more participants exhibiting novelty preference in the motion-defined condition. To perform statistical analyses, we identified participants with novelty preference scores of 55%¹ or higher as “learners”, while identifying the rest as “non-learners”. For Blocks 1–2, in the label-defined condition there were 4 out 21 learners (19%) and in the motion-defined condition there were 7 out 20 learners (35%), which was not statistically different, χ² (1, N = 41) = 1.33, p = .249. However, in Blocks 3–4, there were 13 out 20 learners (65%) in the motion-defined condition, which exceeded the number of learners in the label-defined condition 7 out of 21, (33.3%), χ² (1, N = 41) = 4.11, p = .043. Therefore, despite the fact that there was no advantage for the motion-defined condition during familiarization, there was greater evidence of category learning in this condition than in the label-defined condition.

Individual novelty preference scores in the label-defined and motion-defined conditions in Blocks 1–2 (A) and in blocks 3–4 (B).

Patterns of Attention

We also examined how attention was distributed among different features of the stimuli on test trials (recall that neither labels nor motion was presented during these trials). Since there was no main effect of block (p = .54), the data were collapsed across blocks and results are presented in Figure 5. There was a significant main effect of feature, F(4, 156) = 112.04, MSE = 2.39, p < .001, η_p² = 0.742. Neither the interaction nor the main effect of condition was significant (ps > .117). These results suggest that the difference in novelty preference between the label-defined and the motion-defined conditions did not stem from patterns of attention at test.

Proportion of accumulated looking time to each feature averaged across 4 blocks at test in the label-defined and motion-defined conditions.

Similar analyses were conducted to examine the patterns of attention on familiarization trials. (see Figure 6). The proportion of looking time to each feature was calculated within each trial (8000 ms) and then averaged across five familiarization trials within each block. These data were submitted to a 5 (Feature: head vs. body vs. hands vs. feet vs. antennae) by 4 (Block: 1 vs. 2 vs. 3 vs. 4) by 2 (Condition: label-defined vs. motion-defined) mixed ANOVA, with feature and block as within-subjects factors and condition as a between-subjects factor. Because there was no effect of block (p = .155), data were collapsed across the four blocks. There was a feature by condition interaction, F(4, 156) = 30.95, MSE = 0.391, p < .001, η_p² = 0.442. Infants’ accumulated more looking to the head in the label-defined condition compared to the motion-defined condition, independent sample t(39) = 4.94, p < .001, d = 1.54, whereas they looked significantly longer at the feet in motion-defined condition than in label-defined condition, independent sample t(20.74) = 7.71, p < .001, d = 2.46.

Proportion of accumulated looking time to each feature averaged across 4 blocks at familiarization in the label-defined and motion-defined conditions.

Although greater attention to the feet in the motion-defined condition may not be surprising, it was nevertheless associated with better learning. Note that feet motion was introduced about 2000 ms into the trial and ended about 3000 ms before the end of the trial. Given that participants spent most of the time looking at the head, it is likely that they moved their gaze at least twice: first from the head to the feet and then back from the feet to the head. Therefore, participants in the motion-defined condition were likely to have more gaze shifts, and this more distributed pattern of attention may have led to better category learning.

In order to examine this possibility, we further analyzed the number of fixation shifts between AOIs after the onset of label or motion during familiarization. To perform statistical analyses, we identified a fixation shift as a valid one if the looking time accumulated in the AOI before the shift and in the AOI after the shift were at least 100 ms respectively. The average number of valid fixation shifts within each block across the conditions is presented in Figure 7. These data were submitted to a 4 (Block: 1 vs. 2 vs. 3 vs. 4) by 2 (Condition: label-defined vs. motion-defined) mixed ANOVA, with block as a within-subjects factor and condition as a between-subjects factor. Results revealed a main effect of block, F (3, 117) = 3.92, MSE = 79.22, p = .010, η_p² = 0.091, with the number of valid fixation shifts decreasing through blocks. And more importantly, there was a main effect of condition, F (1, 39) = 5.09, MSE = 442.85, p = .013, η_p² = 0.149, with infants making more shifts after the onset of motion in the motion-defined condition than that after the onset of label in the label-defined condition.

Average number of valid fixation shifts within each block during familiarization in the label-defined and motion-defined conditions.

However, one may argue that greater attention to the head in the label condition (see Figure 6) is indicative of the fact that labels attracted attention to commonalities, with participants optimizing attention by shifting it to the head. Another possibility is that participants in the label-defined condition merely exhibited a head bias (see Quinn et al., 2009, for evidence of head bias in infancy). We therefore deemed it necessary to directly examine these possibilities by examining whether infants exhibited focused attention after the label was introduced and whether infants differentiated the novel head from the familiar head at familiarization.

Dynamics of attention

If labels attract attention to commonalities, attention may be diffused early in the trial, but should become more focused on the head after the label was introduced. However, if attention to the head stems from a bias, attention after the introduction of a label should be no more focused than before. To examine the dynamics of infants’ attention within familiarization trials, attention shifts were calculated every 1000 ms, and then averaged across four blocks at each time point for the total duration of 8000 ms. Data between 1000 ms (i.e., after the word “Look” was introduced in both conditions) and 7000 ms (when infants looking started to decrease rapidly) were used for analysis and were shown in Figure 8. Results presented in Figure 8 indicate that labels did not attract attention to commonalities (i.e., if they did, the number of shifts should have dropped rapidly after the onset of the label). In contrast, there was an increase in the number of shifts after the introduction of motion. Therefore, motion resulted in a more distributed pattern of attention and in better learning. A 6 (Time Point: 1 vs. 2 vs. 3 vs. 4 vs. 5 vs. 6) by 2 (Condition: label vs. motion) mixed ANOVA confirmed these findings. There was a main effect of condition, F (1, 39) = 7.51, MSE = 34.18, p = .009, η_p² = 0.161, with more shifts transpiring in the motion-defined than the label-defined condition. There was also a main effect of time point, F (5, 195) = 22.28, MSE = 13.57, p < .001, η_p² = 0.364, with a quadratic trend showing that shifts increased early in the trial and then decreased later in the trial, F (1, 39) = 54.94, p < .001, η_p² = 0.585. In addition, there was a significant time point by condition interaction, F (5, 195) = 5.07, MSE = 3.09, p < .001, η_p² = 0.115. To examine the interaction, we compared the number of shifts at each time point after the introduction of the label or motion to that before the introduction (i.e., at the first time point). These pair-wise comparisons with Bonferroni correction indicated that in the label-defined condition, the number of shifts at the three consecutive time points after the onset of the label was comparable to that before the label, ps > .203, whereas in the motion-defined condition, there were more shifts at the three consecutive time points after the onset of motion than before, ps < .010. Therefore, there was no evidence that attention became more focused after the label was introduced, but it became more distributed after motion was introduced.

Average number of valid fixation shifts averaged across four blocks at each time point within familiarization trials in the label-defined and motion-defined conditions. Each time point represents a 1000 ms time window.

Looking time to novel vs. familiar head

Another way of examining whether longer looking to the head in the label condition stemmed from a head bias or from increased attention to commonalities, we examined novelty preference during familiarization. Recall that because the categories had the family resemblance structure (see Table 1), each familiarization item had one out-of-category (i.e., novel) feature. Therefore, in each training block, there were four familiarization items with a given head and one familiarization item with a novel head. If labels attract attention to the common head, infants should look longer to the novel head during familiarization. However, if attention to the head stems from a head bias, infants should be interested in the head, whether it is familiar or novel. To examine this, we compared infants’ looking time to the novel head (i.e., the head of item A5 or B5 as shown in Table 1) with the average looking time to the four familiar heads (i.e., the head of items A1–A4 or B1–B4 as shown in Table 1) for each block in the label-defined condition. These data were submitted to a 4 (Block: 1 vs. 2 vs. 3 vs. 4) by 2 (Head Type: novel vs. familiar) within-subjects ANOVA. The results showed that infants did not differentiate the two types of head, with neither the interaction (p = .467) nor the main effect of head type (p = .139) being significant. These results suggest that looking to the head stemmed from a head bias rather than from labels attracting attention to commonalities. Similarly, their looking at other features did not differ when these features were familiar or novel (all ps > .1).

Overall, participants exhibited evidence of category learning in the motion-defined but not in the label-defined condition. This outcome is compatible with two possibilities: (1) label interfering with category learning or (2) motion facilitating category learning. To distinguish between these possibilities, we needed a baseline, in which neither labels nor motion were presented. If participants succeed in the baseline and exhibit patterns of attention comparable to those in the motion-defined condition, then labels interfered with learning. In contrast, if participants fail in the baseline and exhibit patterns of attention comparable to those in the label-defined condition, then motion facilitated learning. The goal of experiment 2 was to distinguish between these possibilities