Discussion 1 PSY545

DOI: 10.1126/science.1175626 , 284 (2009); 325Science

et al.Andrew N. Meltzoff, Foundations for a New Science of Learning

www.sciencemag.org (this information is current as of July 16, 2009 ): The following resources related to this article are available online at

http://www.sciencemag.org/cgi/content/full/325/5938/284 version of this article at:

including high-resolution figures, can be found in the onlineUpdated information and services,

http://www.sciencemag.org/cgi/content/full/325/5938/284#otherarticles , 19 of which can be accessed for free: cites 65 articlesThis article

http://www.sciencemag.org/cgi/collection/education Education

: subject collectionsThis article appears in the following

http://www.sciencemag.org/about/permissions.dtl in whole or in part can be found at: this article

permission to reproduce of this article or about obtaining reprintsInformation about obtaining

registered trademark of AAAS. is aScience2009 by the American Association for the Advancement of Science; all rights reserved. The title

CopyrightAmerican Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by theScience

o n J

u ly

1 6 , 2 0 0

9

w w

w .s

ci e n ce

m a g .o

rg D

o w

n lo

a d e d f ro

m

Foundations for a New Science of Learning Andrew N. Meltzoff,1,2,3* Patricia K. Kuhl,1,3,4 Javier Movellan,5,6 Terrence J. Sejnowski5,6,7,8

Human learning is distinguished by the range and complexity of skills that can be learned and the degree of abstraction that can be achieved compared with those of other species. Homo sapiens is also the only species that has developed formal ways to enhance learning: teachers, schools, and curricula. Human infants have an intense interest in people and their behavior and possess powerful implicit learning mechanisms that are affected by social interaction. Neuroscientists are beginning to understand the brain mechanisms underlying learning and how shared brain systems for perception and action support social learning. Machine learning algorithms are being developed that allow robots and computers to learn autonomously. New insights from many different fields are converging to create a new science of learning that may transform educational practices.

C ultural evolution, which is rare among species and reaches a pinnacle in Homo sapiens, became possible when new forms

of learning evolved under selective pressure in our ancestors. Culture underpins achievements in language, arts, and science that are unprecedented in nature. The origin of human intelligence is still a deep mystery. However, the study of child de- velopment, the plasticity of the human brain, and computational approaches to learning are laying the foundation for a new science of learning that provides insights into the origins of human intelligence.

Human learning and cultural evolution are sup- ported by a paradoxical biological adaptation: We are born immature. Young infants cannot speak, walk, use tools, or take the perspective of others. Immaturity comes at a tremendous cost, both to the newborn, whose brain consumes 60% of its entire energy budget (1), and to the parents. Dur- ing the first year of life, the brain of an infant is teeming with structural activity as neurons grow in size and complexity and trillions of new con- nections are formed between them. The brain continues to grow during childhood and reaches the adult size around puberty. The development of the cerebral cortex has “sensitive periods” during which connections between neurons are more plastic and susceptible to environmental influence: The sensitive periods for sensory processing areas occur early in development, higher cortical areas

mature later, and the prefrontal cortex continues to develop into early adulthood (2).

Yet immaturity has value. Delaying the matu- ration and growth of brain circuits allows initial learning to influence the developing neural archi- tecture in ways that support later, more complex learning. In computer simulations, starting the learn- ing process with a low-resolution sensory system allows more efficient learning than starting with a fully developed sensory system (3).

What characterizes the exuberant learning that occurs during childhood? Three principles are

emerging from cross-disciplinary work in psychol- ogy,neuroscience,machinelearning,andeducation, contributing to a new science of learning (Fig. 1). These principles support learning across a range of areas and ages and are particularly useful in ex- plaining children’s rapid learning in two unique domains of human intelligence: language and so- cial understanding.

Learning is computational. Discoveries in de- velopmental psychology and in machine learning are converging on new computational accounts of learning. Recent findings show that infants and young children possess powerful computational skills that allow them automatically to infer struc- tured models of their environment from the statis- tical patterns they experience. Infants use statistical patterns gleaned from experience to learn about both language and causation. Before they are three, children use frequency distributions to learn which phonetic units distinguish words in their native language (4, 5), use the transitional probabilities between syllables to segment words (6), and use covariation to infer cause-effect relationships in the physical world (7).

Machine learning has the goal of developing computer algorithms and robots that improve automatically from experience (8, 9). For exam- ple, BabyBot, a baby doll instrumented with a video camera, a microphone, and a loudspeaker (10), learned to detect human faces using the temporal contingency between BabyBots’ pro- grammed vocalizations and humans that tended to

1Institute for Learning and Brain Sciences, University of Wash- ington, Seattle, WA 98195, USA. 2Department of Psychology, University of Washington, Seattle, WA, 98195, USA. 3Learning in Informal and Formal Environments (LIFE) Center, University of Washington, Seattle, WA 98195, USA. 4Department of Speech and Hearing Sciences, University of Washington, Seattle, WA, 98195, USA. 5Institute for Neural Computation, University of California at San Diego, La Jolla, CA 92093, USA. 6Temporal Dynamics of Learning Center (TDLC), University of California at San Diego, La Jolla, CA 92093, USA. 7Howard Hughes Medical Institute, Salk Institute for Biological Studies, La Jolla, CA 92037, USA. 8Division of Biological Sciences, University of California at San Diego, La Jolla, CA 92093, USA.

*To whom correspondence should be addressed. E-mail: [email protected]

Fig. 1. The new science of learning has arisen from several disciplines. Researchers in developmental psychology have identified social factors that are essential for learning. Powerful learning algorithms from machine learning have demonstrated that contingencies in the environment are a rich source of infor- mation about social cues. Neuroscientists have found brain systems involved in social interactions and mechanisms for synaptic plasticity that contribute to learning. Classrooms are laboratories for discovering effective teaching practices. [Photo credits: R. Goebel (neuroscience), iStockphoto.com/J. Bryson (education), Y. Tsuno/AFP/Getty Images (machine learning)]

17 JULY 2009 VOL 325 SCIENCE www.sciencemag.org284

REVIEW

o n J

u ly

1 6 , 2 0 0

9

w w

w .s

ci e n ce

m a g .o

rg D

o w

n lo

a d e d f ro

m

respond to these babylike vocalizations. After 6 min of learning, BabyBot detected novel faces and generalized to the schematic faces used in studies of infant face recognition.

Statistical regularities and covariations in the world thus provide a richer source of information than previously thought. Infants’ pickup of this information is implicit; it occurs without parental training and begins before infants can manipulate the physical world or speak their first words. New machine learning programs also succeed with- out direct reinforcement or supervision. Learning from probabilistic input provides an alternative to Skinnerian reinforcement learning and Chomskian nativist accounts (11, 12).

Learning is social. Children do not compute statistics indiscriminately. Social cues highlight what and when to learn. Even young infants are predisposed to attend to people and are motivated to copy the actions they see others do (13). They more readily learn and reenact an event when it is produced by a person than by an inanimate device (14, 15).

Machine learning studies show that system- atically increasing a robot’s social-like behaviors and contingent responsivity elevates young chil- dren’s willingness to connect with and learn from it (16). Animal models may help explain how social interaction affects learning: In birds, neuro- steroids that affect learning modulate brain activity during social interaction (17). Social interaction can extend the sensitive period for learning in birds (18). Social factors also play a role in life- long learning—new social technologies (for ex- ample, text messaging, Facebook, and Twitter) tap humans’ drive for social communication. Edu- cational technology is increasingly embodying the principles of social interaction in intelligent tutor- ing systems to enhance student learning (19).

Learning is supported by brain circuits linking perception and action. Human social and language learning are supported by neural-cognitive systems that link the actions of self and other. Moreover, the brain machinery needed to perceive the world and move our bodies to adapt to the movements of people and objects is complex, requiring contin- uous adaptation and plasticity. Consider what is necessary to explain human imitative learning. Newborns as young as 42 min old match gestures shown to them, including tongue protrusion and

mouth opening (20). This is remarkable because infants cannot see their own faces, and newborns have never seen their reflection in a mirror. Yet, newborns can map from observed behavior to their own matching acts, suggesting shared repre- sentations for the acts of self and others (15, 20). Neuroscientists have discovered a striking over- lap in the brain systems recruited both for the perception and production of actions (21, 22). For example, in human adults there is neuronal activation when observing articulatory movements in the cortical areas responsible for producing those articulations (23). Social learning, imitation, and sensorimotor experience may initially generate, as well as modify and refine, shared neural circuitry for perception and action. The emerging field of social neuroscience is aimed at discovering brain mechanisms supporting close coupling and at- tunement between the self and other, which is the hallmark of seamless social communication and interaction.

Social Learning and Understanding Human children readily learn through social in- teractions with other people. Three social skills are foundational to human development and rare in other animals: imitation, shared attention, and empathic understanding.

Imitation. Learning by observing and imitating experts in the culture is a powerful social learning mechanism. Children imitate a diverse range of acts, including parental mannerisms, speech pat- terns, and the use of instruments to get things done. For example, a toddler may see her father using a telephone or computer keyboard and crawl up on the chair and babble into the receiver or poke the keys. Such behavior is not explicitly trained (it may be discouraged by the parent), and there is no inborn tendency to treat plastic boxes in this way—the child learns by watching others and imitating.

Imitation accelerates learning and multiplies learning opportunities. It is faster than individual discovery and safer than trial-and-error learning. Children can use third-person information (ob- servation of others) to create first-person knowl- edge. This is an accelerator for learning: Instead of having to work out causal relations themselves, children can learn from watching experts. Imitative learning is valuable because the behavioral ac-

tions of others “like me” serve as a proxy for one’s own (15).

Children do not slavishly duplicate what they see but reenact a person’s goals and intentions. For example, suppose an adult tries to pull apart an object but his hand slips off the ends. Even at 18 months of age, infants can use the pattern of unsuccessful attempts to infer the unseen goal of another. They produce the goal that the adult was striving to achieve, not the unsuccessful attempts (14). Children choose whom, when, and what to imitate and seamlessly mix imitation and self- discovery to solve novel problems (24, 25).

Imitation is a challenging computational prob- lem that is being intensively studied in the robotic and machine learning communities (26, 27). It requires algorithms capable of inferring complex sensorimotor mappings that go beyond the repeti- tion of observed movements. The match must be achieved despite the fact that the teacher may be different from the observer in several ways (e.g., size, spatial orientation, morphology, dexterity). The ultimate aim is to build robots that can learn like infants, through observation and imitation (28). Current computational approaches to imi- tation can be divided into direct and goal-based approaches. Direct approaches learn input-action mappings that reproduce the observed behaviors (26); goal-based approaches, which are more re- cent and less explored, infer the goal of the ob- served behaviors and then produce motor plans that achieve those goals (29, 30).

Shared attention. Social learning is facilitated when people share attention. Shared attention to the same object or event provides a common ground for communication and teaching. An early component of shared attention is gaze following (Fig. 2). Infants in the first half year of life look more often in the direction of an adult’s head turn when peripheral targets are in the visual field (31). By 9 months of age, infants interacting with a responsive robot follow its head movements, and the timing and contingencies, not just the visual appearance of the robot, appear to be key (32). It is unclear, however, whether young infants are trying to look at what another is seeing or are simply tracking head movements. By 12 months, sensitivity to the direction and state of the eyes exists, not just sensitivity to the direction of head turning. If a person with eyes open turns to one of

Fig. 2. Gaze following is a mechanism that brings adults and infants into perceptual contact with the same objects and events in the world, facilitating word learning and social communication. After interacting

with an adult (left), a 12-month-old infant sees an adult look at one of two identical objects (middle) and immediately follows her gaze (right).

www.sciencemag.org SCIENCE VOL 325 17 JULY 2009 285

REVIEW

o n J

u ly

1 6 , 2 0 0

9

w w

w .s

ci e n ce

m a g .o

rg D

o w

n lo

a d e d f ro

m

two equidistant objects, 12-month-old infants look at that particular target, but not if the person makes the same head movement with eyes closed (33).

A blindfold covering the person’s eyes causes 12-month-olds to make the mistake of following the head movements. They understand that eye closure, but not a blindfold, blocks the other per- son’s view. Self-experience corrects this error. In a training study, 1-year-olds were given experience with a blindfold so they understood that it made it impossible to see. When the adult subsequently wore the blindfold, infants who had received self- experience with it treated the adult as if she could not see (34), whereas control infants did not. Infants project their own experience onto other people. The ability to interpret the behavior and experience of others by using oneself as a model is a highly effective learning strategy that may be unique to humans and impaired in children with autism (35, 36). It would be useful if this could be exploited in machine learning, and preliminary progress is being made (37).

Empathy and social emotions. The capacity to feel and regulate emotions is critical to under- standing human intelligence and has become an active area of research in human-computer inter- action (38). In humans, many affective processes are uniquely social. Controlled experiments lead to the conclusion that prelinguistic toddlers en- gage in altruistic, instrumental helping (39). Chil- dren also show primitive forms of empathy. When an adult appears to hurt a finger and cry in pain, children under 3 years of age comfort the adult, sometimes offering a bandage or teddy bear (40). Related behavior has been observed with children helping and comforting a social robot that was “crying” (16, 41).

Brain imaging studies in adults show an over- lap in the neural systems activated when people receive a painful stimulus themselves or perceive that another person is in pain (42, 43). These neural reactions are modulated by cultural expe- rience, training, and perceived similarity between self and other (43, 44). Atypical neural patterns have been documented in antisocial adolescents (45). Discovering the origins of individual differ- ences in empathy and compassion is a key issue for developmental social-cognitive neuroscience.

Language Learning Human language acquisition poses a major chal- lenge for theories of learning, and major ad- vances have been made in the last decade (46). No computer has cracked the human speech code and achieved fluent speech understanding across talk- ers, which children master by 3 years of age (11).

Human language acquisition sheds light on the interaction among computational learning, social facilitation of learning, and shared neural circuitry for perception and production.

Behavioral development. Early in development, infants have a capacity to distinguish all sounds across the languages of the world, a capacity shared by nonhuman primates (47). However, infants’ universal capacities narrow with development,

and by one year of age, infants’ ability to perceive sound distinctions used only in foreign languages and not their native environment is weakened. Infants’ universal capacities become language- specific between 9 and 12 months of age. Ameri- can and Japanese infants, who at 7 months of age discriminated /ra/ from /la/ equally well, both change by 11 months: American infants improve significantly while Japanese infants’skills show a sharp decline (48).

This transition in infant perception is strongly influenced by the distributional frequency of sounds contained in ambient language (4, 5). Infants’ com- putational skills are sufficiently robust that labo- ratory exposure to artificial syllables in which the distributional frequencies are experimentally ma- nipulated changes infants’ abilities to discriminate the sounds.

However, experiments also show that the com- putations involved in language learning are“gated” by social processes (49). In foreign-language learn-

ing experiments, social interaction strongly influ- ences infants’statistical learning. Infants exposed to a foreign language at 9 months learn rapidly, but only when experiencing the new language during social interchanges with other humans. American infants exposed in the laboratory to Mandarin Chinese rapidly learned phonemes and words from the foreign language, but only if ex- posed to the new language by a live human being during naturalistic play. Infants exposed to the same auditory input at the same age and for the same duration via television or audiotape showed no learning (50) (Fig. 3). Why infants learned bet- ter from people and what components of social interactivity support language learning are currently being investigated (51). Determining the key stim- ulus and interactive features will be important for theory. Temporal contingencies may be critical (52).

Other evidence that social input advances lan- guage learning comes from studies showing that infants vocally imitate adult vowel sounds by

Fig. 3. The need for social interaction in language acquisition is shown by foreign-language learning experiments. Nine-month-old infants experienced 12 sessions of Mandarin Chinese through (A) natural interaction with a Chinese speaker (left) or the identical linguistic information delivered via television (right) or audiotape (not shown). (B) Natural interaction resulted in significant learning of Mandarin phonemes when compared with a control group who participated in interaction using English (left). No learning occurred from television or audiotaped presentations (middle). Data for age-matched Chinese and American infants learning their native languages are shown for comparison (right). [Adapted from (50) and reprinted with permission.]

17 JULY 2009 VOL 325 SCIENCE www.sciencemag.org286

REVIEW

o n J

u ly

1 6 , 2 0 0

9

w w

w .s

ci e n ce

m a g .o

rg D

o w

n lo

a d e d f ro

m

5 months but not acoustically matched nonspeech sounds that are not perceived as human speech (53, 54). By 10 months, even before speaking words, the imitation of social models results in a change in the types of vocalizations children produce. Children raised in Beijing listening to Mandarin babble by using tonelike pitches char- acteristic of Mandarin, which make them sound distinctly Chinese. Children being raised in Seattle listening to English do not babble by using such tones and sound distinctly American.

Children react to a social audience by increas- ing the complexity of their vocal output. When mothers’ responses to their infants’ vocalizations are controlled experimentally, a mother’s imme- diate social feedback results both in greater num- bers and more mature, adultlike vocalizations from infants (55). Sensory impairments affect infant vocalizations: Children with hearing impairments use a greater preponderance of sounds (such as “ba”) that they can see by following the lip move- ments of the talker. Infants who are blind babble by using a greater proportion of sounds that do not rely on visible articulations (“ga”) (56).

Birdsong provides a neurobiological model of vocal learning that integrates self-generated sen- sorimotor experience and social input. Passerine birds learn conspecific song by listening to and imitating adult birds. Like humans, young birds listen to adult conspecific birds sing during a sen- sitive period in development and then practice that repertoire during a “sub-song” period (akin to bab- bling) until it is crystallized (57). Neural models of birdsong learning can account for this gradual process of successive refinement (58). In birds, as in humans, a social context enhances vocal learn- ing (59).

Neural plasticity. In humans, a sensitive pe- riod exists between birth and 7 years of age when language is learned effortlessly; after puberty, new language learning is more difficult, and native- language levels are rarely achieved (60, 61). In

birds, the duration of the sensitive period is ex- tended in richer social environments (18, 62). Human learning beyond the sensitive period may also benefit from social interaction. Adult foreign- language learning improves under more social learning conditions (63).

A candidate mechanism governing the sensi- tive period for language in humans is neural com- mitment (11). Neural commitment is the formation of neural architecture and circuitry dedicated to the detection of phonetic and prosodic characteristics of the particular ambient language(s) to which the infant is exposed. The neural circuitry maximizes detection of a particular language and, when fully developed, interferes with the acquisition of a new language.

Neural signatures of children’s early language learning can be documented by using event-related potentials (ERPs). Phonetic learning can be docu- mented at 11 months of age; responses to known words, at 14 months; and semantic and syntactic learning, at 2.5 years (64). Early mastery of the sound patterns of one’s native language provides a foundation for later language learning: Children who show enhanced ERP responses to phonemes at 7.5 months show faster advancement in language acquisition between 14 and 30 months of age (65).

Children become both native-language listen- ers and speakers, and brain systems that link per- ception and action may help children achieve parity between the two systems. In adults, func- tional magnetic resonance imaging studies show that watching lip movements appropriate for speech activates the speech motor areas of the brain (66). Early formation of linked perception-production brain systems for speech has been investigated by using brain imaging technology called magneto- encephalography (MEG). MEG reveals nascent neural links between speech perception and pro- duction. At 6 months of age, listening to speech activates higher auditory brain areas (superior temporal), as expected, but also simultaneously

activates Broca’s area, which controls speech production, although listening to non- speech sounds does not [(67); see also (68)]. MEG technol- ogy will allow linguists to ex- plore how social interaction and sensorimotor experience affects the cortical process- ing of language in children and why young children can learn foreign language mate- rial from a human tutor but not a television.

New interactive robots are being designed to teach language to children in a social-likemanner.Engineers created a social robot that autonomously interacts with toddlers, recognizing their moods and activities (16) (Fig. 4). Interaction with the

social robot over a 10-day period resulted in a significant increase in vocabulary in 18- to 24- month-old children compared with the vocabu- lary of an age-matched control group (41). This robotic technology is now being used to test whether children might learn foreign language words through social games with the robot.

Education During their long period of immaturity, human brains are sculpted by implicit social and statis- tical learning. Children progress from relatively helpless, observant newborns to walking, talking, empathetic people who perform everyday exper- iments on cause and effect. Educators are turning to psychology, neuroscience, and machine learn- ing to ask: Can the principles supporting early exuberant and effortless learning be applied to improve education?

Progress is being made in three areas: early intervention programs, learning outside of school, and formal education.

Children are born learning, and how much they learn depends on environmental input, both social and linguistic. Many children entering kinder- garten in the United States are not ready for school (69), and children who start behind in school-entry academic skills tend to stay behind (70). Neuro- science work suggests that differences in learning opportunities before first grade are correlated with neural differences that may affect school learning (71, 72).

The recognition that the right input at the right time has cascading effects led to early interventions for children at risk for poor academic outcomes. Programs enhancing early social interactions and contingencies produce significant long-term im- provements in academic achievement, social ad- justment, and economic success and are highly cost effective (73–75).

The science of learning has also affected the design of interventions with children with dis- abilities. Speech perception requires the ability to perceive changes in the speech signal on the time scale of milliseconds, and neural mechanisms for plasticity in the developing brain are tuned to these signals. Behavioral and brain imaging ex- periments suggest that children with dyslexia have difficulties processing rapid auditory signals; com- puter programs that train the neural systems re- sponsible for such processing are helping children with dyslexia improve language and literacy (76). The temporal “stretching” of acoustic distinctions that these programs use is reminiscent of infant- directed speech (“motherese”) spoken to infants in natural interaction (77). Children with autism spec- trum disorders (ASD) have deficits in imitative learning and gaze following (78–80). This cuts them off from the rich socially mediated learning opportunities available to typicallydeveloping chil- dren, with cascading developmental effects. Young children with ASD prefer an acoustically matched nonspeech signal over motherese, and the degree of preference predicts the degree of severity of their clinical autistic symptoms (81). Children

Fig. 4. A social robot can operate autonomously with children in a preschool setting. In this photo, toddlers play a game with the robot. One long-term goal is to engineer systems that test whether young children can learn a foreign language through interactions with a talking robot.

www.sciencemag.org SCIENCE VOL 325 17 JULY 2009 287

REVIEW

o n J

u ly

1 6 , 2 0 0

9

w w

w .s

ci e n ce

m a g .o

rg D

o w

n lo

a d e d f ro

m

with ASD are attracted to humanoid robots with predictable interactivity, which is beginning to be used in diagnosis and interventions (82).

Elementary and secondary school educators are attempting to harness the intellectual curiosity and avid learning that occurs during natural so- cial interaction. The emerging field of informal learning (83) is based on the idea that informal settings are venues for a significant amount of childhood learning. Children spend nearly 80% of their waking hours outside of school. They learn at home; in community centers; in clubs; through the Internet; at museums, zoos, and aquar- iums; and through digital media and gaming. Informal learning venues are often highly social and offer a form of mentoring, apprenticeship, and participation that maximizes motivation and engages the learner’s sense of identity; learners come to think of themselves as good in tech- nology or as budding scientists, and such self- concepts influence children’s interests, goals, and future choices (84, 85). A recent National Research Council study on science education (83) cataloged factors that enliven learning in informal learning venues with the long-term goal of using them to enhance learning in school.

In formal school settings, research shows that individual face-to-face tutoring is the most ef- fective form of instruction. Students taught by professional tutors one on one show achievement levels that are two standard deviations higher than those of students in conventional instruction (86). New learning technologies are being developed that embody key elements of individual human tutoring while avoiding its extraordinary financial cost. For example, learning researchers have de- veloped intelligent tutoring systems based on cog- nitive psychology that provide an interactive environment with step-by-step feedback, feed- forward instructional hints to the user, and dy- namic problem selection (19). These automatic tutors have been shown to approximate the ben- efits of human tutoring by adapting to the needs of individual students, as good teachers do. Class- rooms are becoming living laboratories as research- ers and educators use technology to track and collect data from individual children and use this information to test theories and design curricula.

Conclusions A convergence of discoveries in psychology, neu- roscience, and machine learning has resulted in principles of human learning that are leading to changes in educational theory and the design of learning environments. Reciprocally, educational practice is leading to the design of new experi- mental work. A key component is the role of “the social” in learning. What makes social interaction such a powerful catalyst for learning? Can key elements be embodied in technology to improve learning? How can we capitalize on social factors to teach children better and to foster their natural curiosity about people and things? These are deep questions at the leading edge of the new science of learning.

References and Notes 1. J. M. Allman, Evolving Brains (Freeman, New York, 1999). 2. S. R. Quartz, T. J. Sejnowski, Behav. Brain Sci. 20, 537

(1997). 3. R. A. Jacobs, M. Dominguez, Neural Comput. 15, 761 (2003). 4. P. K. Kuhl, K. A. Williams, F. Lacerda, K. N. Stevens,

B. Lindblom, Science 255, 606 (1992). 5. J. Maye, J. F. Werker, L. Gerken, Cognition 82, B101 (2002). 6. J. R. Saffran, R. N. Aslin, E. L. Newport, Science 274,

1926 (1996). 7. A. Gopnik et al., Psychol. Rev. 111, 3 (2004). 8. T. M. Mitchell, Machine Learning (McGraw-Hill, New York,

1997). 9. R. Douglas, T. Sejnowski, “Future challenges for the

science and engineering of learning,” www.nsf.gov/sbe/ SLCWorkshopReportjan08.pdf.

10. N. J. Butko, I. R. Fasel, J. R. Movellan, in Proceedings of the 5th IEEE International Conference on Development and Learning, Bloomington, IN, 31 May to 3 June 2006.

11. P. K. Kuhl, Nat. Rev. Neurosci. 5, 831 (2004). 12. A. Gopnik, J. B. Tannenbaum, Dev. Sci. 10, 281 (2007). 13. A. N. Meltzoff, M. K. Moore, Science 198, 75 (1977). 14. A. N. Meltzoff, Dev. Psychol. 31, 838 (1995). 15. A. N. Meltzoff, Acta Psychol. 124, 26 (2007). 16. F. Tanaka, A. Cicourel, J. R. Movellan, Proc. Natl. Acad.

Sci. U.S.A. 104, 17954 (2007). 17. L. Remage-Healey, N. T. Maidment, B. A. Schlinger,

Nat. Neurosci. 11, 1327 (2008). 18. M. S. Brainard, E. I. Knudsen, J. Neurosci. 18, 3929 (1998). 19. K. R. Koedinger, V. Aleven, Educ. Psychol. Rev. 19, 239

(2007). 20. A. N. Meltzoff, M. K. Moore, Early Dev. Parenting 6, 179

(1997). 21. R. Hari, M. Kujala, Physiol. Rev. 89, 453 (2009). 22. G. Rizzolatti, L. Fogassi, V. Gallese, Nat. Rev. Neurosci. 2,

661 (2001). 23. R. Möttönen, J. Järveläinen, M. Sams, R. Hari, Neuroimage

24, 731 (2004). 24. R. A. Williamson, A. N. Meltzoff, E. M. Markman, Dev. Psychol.

44, 275 (2008). 25. B. M. Repacholi, A. N. Meltzoff, B. Olsen, Dev. Psychol.

44, 561 (2008). 26. S. Schaal, Trends Cogn. Sci. 3, 233 (1999). 27. A. P. Shon, J. J. Storz, A. N. Meltzoff, R. P. N. Rao, Int. J.

Hum. Robotics 4, 387 (2007). 28. Y. Demiris, A. N. Meltzoff, Infant Child Dev. 17, 43 (2008). 29. A. Y. Ng, S. Russell, in Proceedings of the 17th International

Conference on Machine Learning, Morgan Kaufmann, Stanford, CA, 29 June to 2 July 2000, pp. 663–670.

30. D. Verma, R. P. N. Rao, in Advances in Neural Information Processing Systems (MIT Press, Cambridge, MA, 2006), pp. 1393–1400.

31. R. Flom, K. Lee, D. Muir, Eds., Gaze-Following (Erlbaum, Mahwah, NJ, 2007).

32. J. R. Movellan, J. S. Watson, in Proceedings of the 2nd IEEE International Conference on Development and Learning, Cambridge, MA, 12 to 15 June 2002, pp. 34–42.

33. R. Brooks, A. N. Meltzoff, Dev. Psychol. 38, 958 (2002). 34. A. N. Meltzoff, R. Brooks, Dev. Psychol. 44, 1257 (2008). 35. A. N. Meltzoff, Dev. Sci. 10, 126 (2007). 36. M. Tomasello, M. Carpenter, J. Call, T. Behne, H. Moll,

Behav. Brain Sci. 28, 675 (2005). 37. J. Bongard, V. Zykov, H. Lipson, Science 314, 1118 (2006). 38. R. W. Picard, Affective Computing (MIT Press, Cambridge,

MA, 1997). 39. F. Warneken, M. Tomasello, Science 311, 1301 (2006). 40. C. Zahn-Waxler, M. Radke-Yarrow, R. A. King, Child Dev.

50, 319 (1979). 41. J. R. Movellan, M. Eckhart, M. Virnes, A. Rodriguez, in

Proceedings of the International Conference on Human Robot Interaction, San Diego, CA, 11 to 13 March 2009, pp. 307–308.

42. T. Singer et al., Science 303, 1157 (2004). 43. G. Hein, T. Singer, Curr. Opin. Neurobiol. 18, 153 (2008). 44. C. Lamm, A. N. Meltzoff, J. Decety, J. Cogn. Neurosci., in

press; available online at www.mitpressjournals.org/doi/ abs/10.1162/jocn.2009.21186.

45. J. Decety, K. J. Michalska, Y. Akitsuki, B. Lahey, Biol. Psychol. 80, 203 (2009).

46. P. Kuhl, L. Gleitman, “Opportunities and challenges for

language learning and education,” www.nsf.gov/sbe/slc/ NSFLanguageWorkshopReport.pdf.

47. P. K. Kuhl, J. D. Miller, Science 190, 69 (1975). 48. P. K. Kuhl et al., Dev. Sci. 9, F13 (2006). 49. P. K. Kuhl, Dev. Sci. 10, 110 (2007). 50. P. K. Kuhl, F.-M. Tsao, H.-M. Liu, Proc. Natl. Acad. Sci. U.S.A.

100, 9096 (2003). 51. B. T. Conboy, P. K. Kuhl, in On Being Moved: From Mirror

Neurons to Empathy, S. Bråten, Ed. (John Benjamins, Philadelphia, 2007), pp. 175–199.

52. P. R. Montague, T. J. Sejnowski, Learn. Mem. 1, 1 (1994). 53. P. K. Kuhl, A. N. Meltzoff, Science 218, 1138 (1982). 54. P. K. Kuhl, A. N. Meltzoff, J. Acoust. Soc. Am. 100, 2425

(1996). 55. M. H. Goldstein, A. P. King, M. J. West, Proc. Natl. Acad. Sci.

U.S.A. 100, 8030 (2003). 56. C. Stoel-Gammon, in Phonological Development,

C. A. Ferguson, L. Menn, C. Stoel-Gammon, Eds. (York, Timonium, MD, 1992), pp. 273–282.

57. M. S. Brainard, A. J. Doupe, Nature 417, 351 (2002). 58. K. Doya, T. J. Sejnowski, in Advances in Neural Information

Processing Systems, G. Tesauro, D. S. Touretzky, T. Leen, Eds. (MIT Press, Cambridge, MA, 1995), pp. 101–108.

59. A. J. Doupe, P. K. Kuhl, in Neuroscience of Birdsong, H. P. Zeigler, P. Marler, Eds. (Cambridge Univ. Press, Cambridge, 2008), pp. 5–31.

60. J. S. Johnson, E. L. Newport, Cognit. Psychol. 21, 60 (1989). 61. R. I. Mayberry, E. Lock, Brain Lang. 87, 369 (2003). 62. L. F. Baptista, L. Petrinovich, Anim. Behav. 34, 1359 (1986). 63. Y. Zhang et al., Neuroimage 46, 226 (2009). 64. P. K. Kuhl, M. Rivera-Gaxiola, Annu. Rev. Neurosci. 31,

511 (2008). 65. P. K. Kuhl et al., Philos. Trans. R. Soc. London Ser. B.

363, 979 (2008). 66. N. Nishitani, R. Hari, Neuron 36, 1211 (2002). 67. T. Imada et al., Neuroreport 17, 957 (2006). 68. G. Dehaene-Lambertz et al., Proc. Natl. Acad. Sci. U.S.A.

103, 14240 (2006). 69. J. P. Shonkoff, D. A. Philips, Eds., From Neurons to

Neighborhoods (National Academy Press, Washington, DC, 2000).

70. G. J. Duncan et al., Dev. Psychol. 43, 1428 (2007). 71. R. D. S. Raizada, T. L. Richards, A. N. Meltzoff, P. K. Kuhl,

Neuroimage 40, 1392 (2008). 72. D. A. Hackman, M. J. Farah, Trends Cogn. Sci. 13, 65 (2009). 73. C. T. Ramey, S. L. Ramey, Merrill Palmer Q. 50, 471 (2004). 74. J. J. Heckman, Science 312, 1900 (2006). 75. E. I. Knudsen, J. J. Heckman, J. L. Cameron, J. P. Shonkoff,

Proc. Natl. Acad. Sci. U.S.A. 103, 10155 (2006). 76. P. Tallal, Nat. Rev. Neurosci. 5, 721 (2004). 77. P. K. Kuhl et al., Science 277, 684 (1997). 78. S. J. Rogers, J. H. G. Williams, Eds., Imitation and the

Social Mind: Autism and Typical Development (Guilford, New York, 2006).

79. P. Mundy, L. Newell, Curr. Dir. Psychol. Sci. 16, 269 (2007). 80. K. Toth, J. Munson, A. N. Meltzoff, G. Dawson, J. Autism

Dev. Disord. 36, 993 (2006). 81. P. K. Kuhl, S. Coffey-Corina, D. Padden, G. Dawson,

Dev. Sci. 8, F1 (2005). 82. B. Scassellati, in Proceedings of the 14th IEEE International

Workshop on Robot and Human Interactive Communication, Nashville, TN, 13 to 15 August 2005, pp. 585–590.

83. P. Bell, B. Lewenstein, A. W. Shouse, M. A. Feder, Eds., Learning Science in Informal Environments (National Academy Press, Washington, DC, 2009).

84. C. D. Lee, Educ. Res. 37, 267 (2008). 85. J. Bruner, The Culture of Education (Harvard Univ. Press,

Cambridge, MA, 1996). 86. B. S. Bloom, Educ. Res. 13, 4 (1984). 87. Supported by NSF Science of Learning Center grants

SBE-0354453 (A.M. and P.K.) and SBE-0542013 (J.M. and T.S.), NIH grants HD-22514 (A.M.) and HD-37954 (P.K.), NSF IIS-0808767 (J.M.), and Howard Hughes Medical Institute (T.S.). The views in this article do not necessarily reflect those of NSF or NIH. We thank M. Asada for assistance with the robotic image in Fig. 1; A. Gopnik, J. Bongard, P. Marshall, S. Cheryan, P. Tallal, and J. Watson for valuable comments; and the members of the NSF Science of Learning Centers for helpful discussions.

10.1126/science.1175626

17 JULY 2009 VOL 325 SCIENCE www.sciencemag.org288

REVIEW

o n J

u ly

1 6 , 2 0 0

9

w w

w .s

ci e n ce

m a g .o

rg D

o w

n lo

a d e d f ro

m