Running Head: Learning mechanism

Can domain-general mechanisms explain the development of biological induction in young children? To answer this question, 43 five-year-olds were presented with a learning (rule discovery) task. The experiment included three between-subject learning conditions and consisted of the Learning and Transfer phases. During the Learning phase, participants received multiple feedback and no-feedback learning trials. On each learning trial, a participant was presented with a Target picture and three Test pictures, such that one of the Test pictures was perceptually identical to the Target, another had the same linguistic label, while the third shared inheritance information. The participant was then told that each of the Test stimuli had a particular biological property (e.g., blue blood vs. yellow blood vs. green blood), and was asked to predict the blood color of the Target. In each of the three learning conditions, participant were taught, by providing them with condition specific “yes/no” feedback, to use either perceptual similarity, linguistic label, or inheritance as a predictor of the biological property in question. Transfer trials using an altered task, took place no less than 60 minutes after the learning phase. It was found that in each learning condition, most participants successfully completed learning and retained this knowledge through the transfer phase, indicating that learning to use predictors could be achieved without increasing domain-specific knowledge. At the same time, it was easier to learn using perceptual similarity as a predictor of a biological property than it was to learn using common label or inheritance information, which points to a pre-existing preference for perceptual similarity as the basis of young children’s induction. Learning mechanism -3 LEARNING TO USE PREDICTORS IN YOUNG CHILDREN’S INDUCTION Induction, or generalization of knowledge from familiar to novel entities or events, is a critical component of human learning and cognition. Those familiar entities, which are used to generalize from, are often referred to as the base of induction, whereas those novel entities, which are used to generalize to, are referred to as the target. Induction is exemplified in the extension of category membership to new members, word meaning to new objects, and properties of a known entity to novel instances, as well as in assigning causes, predicting effects, and formulating general rules. In fact, the centrality of induction for learning made one famous statistician proclaim that induction is "the only process… by which new knowledge comes into the world" (Fisher, 1935). This research is focused on the induction of properties from known entities to novel instances. Examples of such induction are (1) X1 has property Y, therefore X2 has property Y, (2) Xs have property Y, Zs are like Xs therefore Zs have property Y, or (3) X1, X2, …, Xn have property Y, therefore all Xs have property Y. The first type of induction has been defined as specific induction, whereas the latter two have been defined as general induction (see Osherson, Smith, Wilkie, Lopez, & Shafir, 1990, for a discussion). We focus here on specific induction. Simple forms of induction, such as specific induction, seem to be a precondition rather than a product of learning: indeed, before inducing that “knowledge is often generalizable,” one needs to perform a number of knowledge generalizations. Thus, there is little surprise that even infants are able to perform specific induction (Mandler & McDonough, 1996, 1998). For example, in one experiment infants were habituated to a scene of a drinking bird. They later dishabituated to a scene of a drinking airplane, but not to a scene of another drinking bird, thus indicating that Learning mechanism -4 they had generalized the ability to drink from one bird to another (Mandler & McDonough, 1996). Furthermore, 3-4 year-old children were shown to exhibit a great deal of inductive sophistication, distinguishing properties that could be generalized from those that cannot (Gelman & Markman, 1986; Gelman, 1988; Gutheil, Vera, & Keil, 1998; Springer, 1992). For example, children have exhibited knowledge that the property “has thick bones” could be generalized from one bird to another, whereas “has a missing feather” cannot. Finally, previous research found that children do use a variety of information cues when performing induction. These cues include (but are not limited to) labeling information (Gelman & Markman, 1986; 1987; Sloutsky & Lo, 2000b), kinship information (Gutheil, Vera, & Keil, 1998; Johnson & Solomon, 1997; Springer, 1992; 1996; Springer & Keil, 1989; Taylor, 1996), and appearance (Gelman & Markman, 1987; Sloutsky & Lo, 2000b; see also Gelman & Medin, 1993; Keil, 1989; Smith & Jones, 1993, for reviews and discussions). For example, when compared animals either share a label, belong to the same kin, or look alike, young children are more likely to induce a biological property from one animal to the other than when they do not share any of these properties. While it appears early in development, inductive inference does undergo important developmental changes. First, several researchers demonstrated that young children are not sensitive to statistical properties of the base sample, such as sample size and sample diversity (Gutheil & Gelman, 1997; Lopez, Gelman, Gutheil, & Smith, 1992). In particular, young children are equally likely to generalize a biological property from Test stimuli to the Target when the Test stimulus included a single animal or when it included multiple animals (Gutheil & Gelman, 1997; Lopez, et al., 1992; Sloutsky & Lo, 2000a). At the same time, adults are more Learning mechanism -5 likely to generalize properties from multiple examples (Gutheil & Gelman, 1997; Nisbett, Krantz, Jepson, and Kunda, 1983; Osherson, et al., 1990), indicating that they are sensitive to sample size. In addition, young children did not consider induction from several diverse examples to be stronger than induction from several homogenous examples, whereas adults are sensitive to sample diversity and they consider the former induction to be stronger than the latter (Gutheil & Gelman, 1997; Osherson, et al., 1990). For example, adults, but not children, considered the inference that lions have biological property X to be stronger when told that rabbits and elephants have the property (i.e., a diverse base) than when told that rabbits and hares have the property (i.e., a homogeneous base). Previous research also examined relative contributions of various informational cues to inductive inference by pitting these cues against each other, and developmental changes in these contributions (e.g., Gelman & Markman, 1986; Sloutsky & Lo, 2000a; Springer, 1992). In a typical task examining the relative contribution of information cues, a child is presented with a triad of pictures, two of which are Test items, and one is the Target item. Test 1 has a particular appearance A1 and a particular category label L1, while Test 2 has another appearance A2 and a different category label L2. Finally, the Target shares appearance with Test 1 and the category label with Test 2 (i.e., A1L2), and the child is asked to generalize a biological property from one of the Test items to the Target. In a variation of this task, instead of a shared label, shared inheritance has been introduced (Springer, 1992). This research showed that children are more likely to rely on category label than on appearance (Gelman & Markman, 1986) and on inheritance information than on appearance (Springer, 1992). In addition, in the course of development, attentional weights of different informational cues changed with weights of Learning mechanism -6 inheritance increasing and weights of appearance decreasing (Sloutsky & Lo, 2000c). In particular, preadolescents are more likely than young children to rely on inheritance information. Finally, some researchers examined joined contributions of information cues by using stimuli where cues are bundled together (e.g., Sloutsky & Lo, 2000b, 2000c). To examine the joint contribution, children were presented with tasks where one Test stimulus shared several cues with the Target, whereas another Test stimulus shared a single cue. For example, Test 1 shared inheritance and appearance with the Target, whereas Test 2 shared the label with the Target. It was found that young children were more likely to generalize biological properties from a Test stimulus that shared several information cues with the Target than from the one that shared only one information cue, regardless of the type of cues (Sloutsky & Lo, 2000b, 2000c). At the same time, regardless of the number of shared cues, preadolescents and adults invariably relied on inheritance as the most predictive source of information. It was concluded, therefore, that young children integrated multiple cues (or features) when performing inductive inference, whereas preadolescents and adults relied on a single source of information that they deemed most predictive. These findings are consistent with other research indicating that processing develops from holistic to dimensional (Shepp, 1978; Shepp & Swartz, 1976; Smith, 1989a, 1989b). In short, previous research found that (a) young children are not sensitive to statistical properties of the base sample, such as sample size and sample diversity and (b) linguistic labels and kinship information more readily drive young children’s induction than perceptual similarity. It was also found that induction undergoes important developmental changes: (1) young children use multiple sources of information, while preadolescents and adults rely on a single most predictive source, and (2) preadolescents are more likely to rely on inheritance information in induction of biological properties than young children. Learning mechanism -7 These developmental transitions further indicate that in the course of development, weights of different information cues change, with weights of some cues (e.g., inheritance) increasing and other cues (e.g., similar appearance) decreasing. However, while changes themselves are documented (Sloutsky & Lo, 2000c), reasons for these changes remain unclear. The goal of current research is to examine these reasons. Theoretically, this transition from integration of multiple features to induction based on a single feature could stem from a variety of sources. First, this transition could stem from the increase in the domain-specific biological knowledge: preadolescents learn in biology class that inheritance is a causal determiner of anatomical and physiological properties, whereas a common label or similar appearance are not causal determiners. Alternatively, the transition could stem from a domain-general mechanism of probabilistic learning. In this case, even if they do not know causal connections between inheritance and biological properties, common inheritance could be more strongly associated with common biological properties than common label or common appearance, because the former co-occur more frequently than the latter. As children grow, they accumulate more evidence supporting these associations, and, as a result, these associations get stronger. Finally, the transition could stem from a combination of the domainspecific knowledge of the causal importance of inheritance and domain-general knowledge of conditional probabilities or contingencies (cf. Cheng, 1997). In this research, we specifically focus on the role of domain-general factors in the development of induction. In particular, we attempt to train young children to rely on a single cue (predictor) without explaining the causal importance of this predictor for inducing biological properties (i.e., without introducing domain-specific knowledge). Recall that in prior research young children were shown to rely on multiple cues (as opposed to a single cue) when the cues Learning mechanism -8 were bundled, or to rely on common label or inheritance when these cues were pitted against appearance (Sloutsky & Lo, 2000b, 2000c). Note that inheritance, linguistic labels, and perceptual similarity have different causal status with respect to induced biological properties. While inheritance could be construed as a direct cause of biological properties, appearance cannot be construed as a cause of biological properties. At the same time, linguistic labels could be construed either as a marker of an essence, which in turn causes biological properties, or as cue co-varying with biological properties in a non-causal manner. Furthermore, unlike inheritance, which is specific to the domain of biology, common appearance and common labels frequently co-vary with hidden properties (e.g., an object’s function) across various domains. If, regardless of the causal status of the predictor, learning to use each of these cues as predictors would be successful, this would indicate that knowledge of causal relations between the predictor and predicted property is not necessary for successful induction. Successful learning would also indicate that young children have requisite attentional capacities to focus on a learned information cue, while ignoring the rest. Of course, because domain-specific knowledge is not actively manipulated in current research, successful learning would be unable to either confirm or eliminate possible contributions of domain-specific knowledge in the development of inductive inference. In this research, we used a paradigm that made use of both separate and bundled information cues. In the Learning phase, three cues (i.e., inheritance information, label, and appearance) were pitted against each other, and participants were trained to base their induction on a single cue through the use of “yes/no” feedback. In the transfer phase, two information cues (e.g., appearance and label) were bundled together and pitted against the learned cue (e.g., Learning mechanism -9 inheritance). Participants were presented with both implicit and explicit feedback learning trials and with no-feedback transfer trials. This paradigm affords answers to several important questions. First, what are attentional weights of different information cues when these cues are non-bundled and pitted against each other? Second, can these weights be readjusted in the course of learning without an increase in domain-specific knowledge? Third, do potential causal links between predictors and induced biological properties support learning? And fourth, can feature-integration be overridden, that is, can reliance on a single source of information observed in preadolescents and adults be learned by young children? To answer the posed questions, we analyzed learning across the three types of cues, focussing on the number of trials to criterion and the quality of transfer. Answers to the posed questions are critically important for understanding mechanisms of knowledge acquisition as well as changes in these mechanisms in the course of development and learning. These answers are also important for better understanding of the nature of young children’s induction, which could be a function of knowledge-independent attentional mechanisms (Sloutsky & Lo, 2000b; Smith & Jones, 1993), knowledge-dependent naïve theories (Gelman & Coley, 1991; Gelman & Medin, 1993; Gelman, Coley, & Gottfried, 1994), or a combination of both. EXPERIMENT The experiment included three between-subject learning conditions and consisted of the Learning and Transfer phases. During the Learning phase, each participant received up to 10 learning trials. On each learning trial, a participant was presented with a Target picture and three Test pictures of imaginary animals, such that one of the Test animals was perceptually identical to the Target, another had the same linguistic label, while the third shared inheritance Learning mechanism -10 information. After that, the participant was told that each of the Test stimuli had a particular biological property (e.g., blue blood vs. yellow blood vs. green blood), and was asked to predict the blood color of the Target. In each of the three learning conditions, participants were given different feedback. In the Inheritance condition, they were given positive feedback if they induced properties from the Test sharing inheritance and negative feedback if they induced properties from Tests sharing other cues. At the same time, in the Label or Perceptual Similarity conditions they were given positive feedback if they induced along the same label or the same appearance respectively but negative feedback for inducing properties from Tests sharing other cues. Note that on each trial, participants were presented with different pictures, labels, and biological properties. Therefore, within each learning condition, the task of learning was to formulate a general rule (e.g., if a Test and the Target share the label, they also share biological properties) from a number of specific observations (e.g., Two items labeled “a Pofa” had green blood). Participants Forty-three preschool children (mean age = 5.0; SD = 0.29 years; 20 boys and 23 girls) participated in this experiment. These participants were recruited from daycare centers located in upper middle class suburbs of Columbus, Ohio, and were selected on the basis of returned permission slips. Design The experiment had a mixed design with Learning condition as a between-subject factor, and Experimental phase as a within-subject variable. The Learning condition had three levels: (1) Label (L), (2) Perceptual Similarity (PS), and (3) Inheritance (INH), with participants being randomly assigned (stratifying by gender) to one of these conditions. During learning, Learning mechanism -11 participants were taught to rely on one information cue (i.e., Label, Perceptual Similarity, or Inheritance) for inducing biological properties. The experiment included two within-subject Experimental phases: (1) Learning and (2) Transfer. Materials Materials consisted of line-drawing pictures, labels, and biological properties. The experimenters developed 120 line-drawing pictures of animals (4 stimuli * 10 trials * 3 conditions = 120 stimuli) for the Learning Phase; 108 of these pictures were used in the Transfer phase (3 stimuli * 12 trials * 3 conditions = 108 stimuli). The pictures were constructed so as not to resemble any actually existing animals. To further avoid confounds with existing knowledge about specific animals, artificial labels were used. These auditorily presented labels consisted of short two-syllable words presented as count nouns (e.g., a Guga, a Pofi, a Boto, etc.). The biological properties to be generalized from one of the Test stimuli to the Target referred to different colors of blood and bones (e.g., red bones vs. brown bones vs. white bones). Procedure Each child was tested individually in a room outside the classroom. The Learning and Transfer phases were conducted on the same day; the Transfer phase was begun no less than 60 minutes after completing the Learning Phase. This between-phase interval was adopted due to the amount of time required to complete each Phase (approximately 20 minutes) and so as not to overwhelm the children; children returned to their classrooms during this interval. The Experiment was administered on a Dell Inspiron 3500 laptop computer using the program Superlab 2.0 (Cedrus Corporation, 1999). Learning Phase. The Learning phase consisted of a maximum of 10 learning trials with each trial using four stimuli (a Target stimulus and three Test stimuli). Each of the Test stimuli shared Learning mechanism -12 one information cue with the Target, while differing on the other two cues (see Figure 1). As shown in Figure 1, one of the Test stimuli looks like the Target, another shares a label with the Target, whereas the third shares inheritance information with the Target. Each trial used different labels, pictures, and biological properties. The positioning of the Test stimuli relative to the target was varied across trials. Participants were asked to generalize an unobservable biological property (e.g., blood color) from one of the Test stimuli to the Target. After making the generalization, participants were presented with one of three types of feedback—implicit feedback, explicit feedback or no feedback (see Figure 2). As shown in Figure 2, on trials 1 to 5, participants’ responses were followed by implicit feedback (either “That’s right!” or “Try again.”). “That’s right!” feedback was provided only when a participant made a conditionconsistent response (e.g., induction based on a shared label in the Label Condition), otherwise feedback “Try again” was provided. On Trials 6 and 7, participants received explicit feedback. After making a response on these trials, participants were told whether or not their response was correct (i.e., condition consistent) and what was the correct response. For example when inducing on the basis of label in the label learning condition, a participant would be told, “That’s right, the correct one is the one with the same name.” On the other hand, when inducing on the basis of another cue in the label learning condition, a participant would be told, “Try again.” On Trials 8 through 10, participants were presented with no feedback after making their generalizations. On each of the implicit and explicit feedback trials, if a participant responded in a manner inconsistent with a respective learning condition (e.g., a participant in the Inheritance condition basing induction on perceptual similarity), the Test stimulus that led to an erroneous response was removed. The participant was then asked to perform induction with the two remaining Test Learning mechanism -13 stimuli. If the participant again responded in a manner inconsistent with a respective learning condition, the Test stimulus that led to an erroneous response was removed again, and the participant was then asked to perform induction with the one remaining Test stimulus. In the latter case, the participant had no other choice but to perform induction in a manner consistent with the respective learning condition. After making a condition-consistent response, participants received either implicit or explicit feedback and proceeded to the next trial. When condition-consistent responses were given on three consecutive learning trials, the Learning Phase was discontinued. If three consecutive condition-consistent responses were not obtained by Trial 5, then on Trials 6 and 7 participants were given explicit feedback. Those participants who achieved three consecutive condition-consistent responses by Trial 10 proceeded to the Transfer phase. Those participants who did not achieve that criterion did not participate in the Transfer phase. Note that more than 85% of participants in each condition did achieve the learning criterion. Transfer Phase. Participants who made three consecutive condition-consistent responses by Trial 10 of the Learning Phase participated in the Transfer phase that started no less than 60 minutes after completing the Learning Phase. The Transfer phase consisted of 12 trials. On each of these 12 trials, participants were presented with three stimuli (a Target and two Test stimuli). On some trials, one Test stimulus shared a learned property with the Target (e.g., the label), whereas another Test stimulus shared one of the remaining properties (e.g., appearance) with the Target. We will refer to these trials as T-L-1, with T referring to the Target, L referring to a learned cue, and 1 referring to another property pitted against the learned one. On other trials, one Test stimulus shared a learned property with the Target (e.g., the label), whereas another Test stimulus shared both of the remaining properties (e.g., appearance and inheritance) with the Learning mechanism -14 Target. We will refer to these trials as T-L-2, with T referring to the Target, L referring to a learned cue, and 2 referring to two properties pitted against the learned one. Each participant received 8 T-L-1 transfer trials (4 trials for each of the two remaining properties), and 4 T-L-2 transfer trials. An example of the T-L-2 combination in the Perceptual Similarity Condition (PS vs. L + Inh) is given in Figure 3. After completing the experiment, each participant was debriefed. They were reminded that this was just a game and were told that in real life animals are like the ones that gave birth to them. Memory Check. On Trial 6 of the Learning Phase and on Trial 11 of the Transfer Phase a memory check was conducted to ensure that children remembered the cues associated with each Test. The participants were asked to indicate which stimuli shared a label, shared inheritance information, and possessed a certain biological property. No feedback as to the accuracy of the response was provided. Results Results of the Memory Check indicated that the majority of the children were able to remember the stimuli that shared a label, shared inheritance information, and possessed a certain biological property on at least one of the two memory check trials. Of those who participated in the Transfer Phase, 83% passed the memory check for labels, 94% passed for inheritance, and 97% passed for biological properties. All the two children who did not pass the memory check for labels were in the PS learning Condition. In fact, all the children in the Label Condition passed the memory check for labels, and all the children in the Inheritance Condition passed the memory check for inheritance. In what follows, we first analyze participants’ performance in the Learning Phase, followed by analyses of their performance in the Transfer phase. First, we deemed it necessary to examine Learning mechanism -15 initial feature preferences by analyzing participants’ first and second choices on the very first trial, i.e., prior to any effects of training. Results indicate that on the first learning trial across all three conditions, approximately 81% of all children chose the stimulus that was perceptually similar (PS) to the Target as their first induction choice (above chance, Confidence Interval from 63% to 92%, p < .05). Given PS as a first choice, 67% (not different from chance) of the children used L as a second choice while 33% (not different from chance) chose INH as a second choice. In other words, before training, children had a strong preference for using PS to perform inductions with L and INH being chosen at chance levels as the second choice. The second goal of our analysis was to examine whether this initial preference can be overridden and whether the ease of learning depends on the Learning Condition. To achieve this goal, we compared learning indicators, such as the number of learning trials to criterion and the number of feedback trials, across the Learning Conditions (see Table 1). The number of learning trials to criterion was subjected to a one-way ANOVA with Learning Condition as a factor. There was a main effect of Learning Condition, F(2, 35) = 59.82, MSE = 1.66, p <. 0001. Post hoc Bonferroni tests pointed to the following differences in the number of learning trials across the Learning Conditions: L > INH > PS, all ps < .01. The analysis indicated that more trials were needed to acquire Label as a predictor than the other cues, and the least number of trials were needed to acquire PS as a predictor, with inheritance falling in between. These findings were further supported in a related analysis of the mean number of feedback trials required to reach the learning criterion. These numbers were also subjected to a one-way ANOVA with Learning Condition as a factor. Similar to the previous analysis, there was also a significant main effect of Learning Condition, F(2, 40) = 58.46, MSE = 0.24, p < .0001. Again, Learning mechanism -16 post hoc Bonferroni tests for multiple comparisons indicated significant differences among all conditions: L > INH > PS, all ps < .05. These findings indicated differences in how easily children learned to respond in a conditionconsistent manner depending upon condition. As previously stated, children had a strong preference for using PS as their first choice on Learning Phase trial 1; as shown here, children were quickest to reach criterion and consequently required fewer feedback trials when they were in the PS condition. While preferences indicated that as a second choice L and INH were chosen at statistically equivalent frequencies (with a slight preference for L), the L condition required the greatest number of learning trials to reach criterion as well as the greatest number of feedback trials (see Table 1). To compare the speed of learning across the conditions, we also analyzed the number of choices consistent with the respective learning condition across learning trials. These data are presented in Figure 4. Data in the figure indicate that while in the PS condition, participants reliably answered in the condition-consistent manner from the outset of learning; in the L and INH conditions they were at chance until receiving an explicit feedback. Data in Figure 4 also indicate that the initial preference for perceptual similarity can be overridden: In all three conditions, at the end of learning, the number of condition consistent responses was significantly