Gopnik-CH_02.qxd 12/25/2006 4:44 PM Page 37 Meltzoff, A. N. (2007). Infants’ causal learning: Intervention, observation, imitation. In A. Gopnik & L. Schulz (Eds.), Causal learning: Psychology, philosophy, and computation (pp. 37-47). Oxford: Oxford University Press. 2 Infants’ Causal Learning Intervention, Observation, Imitation Andrew N. Meltzoff Infants’ Understanding of Interventions covariation to be understood as fully causal. The by Self and Other concept of an intervention may help us move beyond a debate about the primacy of perception Causal learning by children combines both observa- (Michotte) versus action (Piaget) to theories that map tion and action. These two sources of information observations and actions to the same abstract causal have not been well integrated in developmental representations. theory. Following Michotte (1963), some develop- For developmental scientists, one striking feature mental scientists argue that young infants are exquis- of the philosophical notion of an intervention is that itely tuned observers, and that their perceptual it is abstract—an intervention can be performed by understanding of causality far outstrips their ability to the self or by another person (or even by a “natural use this information to manipulate the world. experiment” not involving an agent). We can learn Following Piaget (1954), others argue that young not only through our own interventions on the world, infants learn little by pure observation—self-produced but also by watching the interventions of others. This motor action is critical; cognitive development gener- intriguing idea is incompatible with many classical ally, and causal reasoning in particular, is charted as a views of infancy, which explicitly deny the equiva- progressive combination of action schemes. lence between observing others and acting oneself. Bayes net approaches provide a way of using both In classical developmental views, we observe others observation and action (in the form of “interven- from the outside as a series of movements in space, tions”), combining them to generate veridical repre- but we feel ourselves from the inside as yearnings, sentations of the causal structure in the world. In intentions, and freely willed plans. The way we repre- fact, on some interpretations (Woodward, 2003), the sent self versus other is fundamentally different. This link to intervention is crucial for observed patterns of results in a disconnect between learning by doing 37 Gopnik-CH_02.qxd 12/25/2006 4:44 PM Page 38 38 CAUSATION AND INTERVENTION (self-action) and learning by watching (other’s action). and an observer Pigeon 2 (P-2) is reinforced for peck- A prime developmental achievement is to bring these ing on seeing this event, then P-2 will eventually be two modes of learning into line. shaped to peck when seeing P-1 pecking. But, P-2 did There are many ways of testing the psychological not learn this intervention on the basis of observing linkage between observed and executed interventions. the other animal. All that has happened is that the I have used infant imitation, which has several behavior of P-1 has become a cue for eliciting a con- virtues. First, imitation is natural to humans, even ditioned response in P-2. It follows that the observer babies. Second, in imitating novel acts, infants fashion pigeon could be conditioned to perform a nonimita- their interventions based on observing interventions tive act just as easily. Skinner (1953) endorses this performed by others. Third, it is widely acknowledged implication: “The similarity of stimulus and response that humans are far more proficient imitators than in imitation has no special function. We could easily other primates (Meltzoff, 1996; Povinelli, 2000; establish behavior in which the ‘imitator’ does exactly Tomasello & Call, 1997), and therefore we may be the opposite of the ‘imitatee’ ” (p. 121). getting at distinctively human cognition by examining It is known that human infants as young as 3 to 6 human imitation and its development. Fourth, com- months old can be operantly conditioned quite readily putational models including Bayesian approaches have (e.g., Rovee-Collier, 1990). This means that they can been applied successfully to both human and robotic learn the contingency between their own actions and imitation (e.g., Demiris & Meltzoff, in press; Meltzoff & results in the world. But, the capacity for operant condi- Moore, 1983; Rao, Shon, & Meltzoff, 2007). tioning does not mean that the infant can learn these Historically, there are two principal theories of action-outcome relations from observing the acts of oth- how infants come to imitate the acts of others: ers. In other words, the fact that infants can learn an Skinnerian and Piagetian theory. I argue that neither intervention through their own trial and error (learning of these can encompass the modern empirical work by doing) does not mean that they can learn to perform on infant imitation. The new data are more compati- the intervention on the basis of observing the interven- ble with the view that there is a fundamental equiva- tions of others (learning by watching). The latter would lence between the perception and performance of be imitation. The former is just a special case of operant goal-directed acts—an abstract mapping connecting conditioning in which a friendly demon (a clever acts seen and acts done—that was not envisioned in mother or experimenter) has arranged it so the discrim- the classical frameworks. inative cue matches the reinforced response. The moral Skinner (1953) proposed that young infants cannot is that if we want to know whether infants can learn an imitate the acts of others without specific training. intervention through observation, then we need to When a young infant sees a mother perform an act know the infant’s reinforcement history or, failing that, such as shaking a rattle to make a sound, the infant use a novel act for which prior shaping is unlikely. does not know what movements to recruit to copy this Piagetian theory (1962) came to similar conclu- act. Rather, the mother needs to shape the child’s sions as Skinner, albeit for entirely different reasons. response through operant conditioning. Mom shakes Piaget also thought that young infants could not imi- the rattle, and then the infant responds with random tate spontaneously. In Piaget’s case, it was not that motor acts. Mom selectively reinforces those acts that infants needed to be conditioned to learn to imitate, are similar to shaking the rattle. Over time, the but rather that they needed to reach a certain stage of mother’s shaking comes to serve as a discriminitive cue cognitive sophistication. Piaget realized that translat- (a bell or a light would do as well) that elicits the rein- ing a seen intervention into one executed by the self forced act (the baby’s rattle shaking). To the outside was nontrivial, and he claimed it was beyond the observer, the infant is imitating, but this is not because capacity of infants in the first half year of life. He the baby is able to translate the acts seen into acts hypothesized that infants were “egocentric,” even done. The parent essentially teaches the infant what to “solipsistic.” The youngest infants could not learn do and when to do it through operant conditioning. novel acts from observing others (whether these acts This is not an entirely hypothetical example. In were complex means-ends relationships or simple fact, Skinner (1953) has shown that pigeons can be body acts) because learning at first occurred through conditioned to peck a key when they see other pigeons self-action independently of other people (what peck: If Pigeon 1 (P-1) pecks at a key to obtain food Piaget called practical intelligence). Gopnik-CH_02.qxd 12/25/2006 4:44 PM Page 39 INFANTS’ CAUSAL LEARNING 39 The Piagetian concept of infantile egocentrism was body scheme that maps from observed body parts to most famously illustrated in his predictions about facial their own body, despite never having seen their own imitation. Infants can see you make a facial movement, face. Similarly, they responded accurately to lip but they cannot see their own faces. If the infant is protrusion versus lip opening, showing that different young enough, he or she will never have seen his or her patterns of action can be extracted and imitated when face in a mirror. How could the infant link the observed the specific body part is controlled. facial acts of others with personal unseen bodily acts? As my psychology colleagues quickly pointed out, According to Piaget, this “invisible imitation” was these infants may not have been young enough to impossible because self and other were known in such answer the objections of Skinner and Piaget. In their different terms; there was no abstract framework for 2 weeks of life, they might have learned the relevant connecting observation and performance. Piaget associations. Perhaps mothers conditioned their chil- (1962) put it this way: “The intellectual mechanism of dren to stick out their tongues whenever they saw this the child will not allow him to imitate movements he gesture. The definitive test involved newborns who sees made by others when the corresponding move- averaged 32 hours old at the time of the test. The old- ments of his own body are known to him only tactually est infant was 72 hours old, and the youngest was just or kinesthetically (as, for instance, putting out his 42 minutes old. The newborns accurately imitated tongue) . since the child cannot see his own face, (Meltzoff & Moore, 1983, 1989). Apparently, facial there will be no imitation of movements of the face imitation is innate. This suggests a fundamental [before approximately 1 year old]” (p. 19). equivalence between the perception and production Thus, Piaget shared Skinner’s view that actions of acts that is built into the mind of the human baby.