Visual statistical learning in infancy: evidence for a domain general learning mechanism

Abstract

The rapidity with which infants come to understand language and events in their surroundings has prompted speculation concerning innate knowledge structures that guide language acquisition and object knowledge. Recently, however, evidence has emerged that by 8 months, infants can extract statistical patterns in auditory input that are based on transitional probabilities defining the sequencing of the input's components (Science 274 (1996) 1926). This finding suggests powerful learning mechanisms that are functional in infancy, and raises questions about the domain generality of such mechanisms. We habituated 2-, 5-, and 8-month-old infants to sequences of discrete visual stimuli whose ordering followed a statistically predictable pattern. The infants subsequently viewed the familiar pattern alternating with a novel sequence of identical stimulus components, and exhibited significantly greater interest in the novel sequence at all ages. These results provide support for the likelihood of domain general statistical learning in infancy, and imply that mechanisms designed to detect structure inherent in the environment may play an important role in cognitive development.

Introduction

A central question asked by developmental psychologists concerns how infants learn so much in so little time, often with little explicit instruction. The rapidity and ease with which children understand and produce speech, for example, have led to the postulation of an innate device that allows the young child to discover how his or her native language embodies those principles common to all languages (Chomsky, 1965, Gleitman and Wanner, 1982, Pinker, 1984). In object perception tasks, likewise, infants' facility at recognizing apparent violations of physical laws (e.g. the persistence of objects after occlusion) has been cited as evidence for innate reasoning skills that provide some knowledge about fundamental object properties (Carey, 1999, Spelke and Van de Walle, 1993, Wynn, 1998). On such accounts the role of learning is subsidiary to a set of specialized cognitive modules, operational early in infancy, from which domain specific knowledge emerges (cf. Fodor, 1983). Such views may risk neglecting the potential role of environmental structure in guiding development. Natural visual scenes, for example, are richly structured and contain a high degree of statistical redundancy, showing considerable consistency across space and time (Field, 1994), and there is evidence that development of response properties of some visual neurons exploits the statistical nature of the input (Olshausen & Field, 1996). In the present study, we asked whether young infants are able to extract statistical information in visual stimuli, to explore a larger question of the nature of learning in infancy: are early learning mechanisms attuned to input structure in the visual environment?

Several recent experiments have reported evidence that infants readily learn statistically defined patterns in auditory input sequences. Saffran, Aslin, and Newport (1996), for example, investigated 8-month-old infants' word segmentation in a corpus of artificial speech. They noted that in natural speech, adjacent sounds that co-occur with a high probability are usually found within words, whereas low probability sound pairs tend to span word boundaries. This difference in likelihood of co-occurrence provides potential information for word boundaries, and may contribute to early language acquisition by bolstering the ability to segment the speech stream into meaningful units.

To investigate infants' discrimination of high and low probability sound pairs within a corpus of speech, Saffran et al. (1996) presented 8-month-olds with a synthesized speech stream consisting of four three-syllable “words” composed of 12 unique syllables (e.g. tupiro, golabu, dapiku, and tilado), presented in random order (e.g. dapikutupirotiladogolabutupiro…) for 2 min. The only cues to word boundaries were the transitional probabilities between syllable pairs. For example, the transitional probability of tu-pi in this corpus was 1.00, because pi always followed tu within the word tupiro (a within-word syllable pair), whereas the transitional probability of ro-go was 0.33, because golabu was one of three words that could follow tupiro (a between-word syllable pair). Following exposure, the infants then heard isolated instances of familiar words, presented repeatedly. In one test condition, the familiar words were presented randomly within blocks that also contained “non-words”, and in another test condition the familiar words were presented with “part-words”. Non-words were created from novel combinations of syllables from the familiar speech stream (e.g. pitugo), and part-words were created by combining the last syllable of one word with the first two syllables of a second word (e.g. kutila). The infants showed a reliably greater interest in the non-words and part-words than in the words, as revealed by a preferential head-turning paradigm. On the logic that infants often exhibit post-familiarization novelty preferences (Bornstein, 1985), this suggests that they distinguished between the words and the other stimuli based on learning the transitional probabilities defining word boundaries. (See Aslin, Saffran, and Newport (1998) for evidence that the results stem from true computation of input statistics rather than simple frequency counting.)

These findings provide evidence for a functional statistical learning mechanism, available to infants, that segments speech into words on the basis of computation of input statistics, and gives rise to questions concerning the generality of this ability. Saffran, Johnson, Aslin, and Newport (1999) found that 8-month-olds detected transitional probabilities of non-linguistic tone sequences, indicating that statistical learning is not a purely linguistic mechanism. Further evidence for generality comes from experiments by Hauser, Newport, and Aslin (2001) with non-human primates (cotton-top tamarins, a species of New World monkey). After exposure to the same set of auditory stimuli employed by Saffran et al. (1996), adult monkeys showed reliably greater interest in both non-words and part-words than in the familiar words, suggesting that they were able to extract the statistical information defining word boundaries in the artificial speech, in like manner to human infants. These experiments imply that statistical learning may be a general purpose learning device, but it is unknown at present if young infants can detect statistically defined structure in sequential visual stimuli.

The present study had two goals. First, we probed the question of domain generality of statistical learning in infancy by asking whether it is limited to auditory information. This was accomplished with a task in which infants were presented with sequential visual input that contained probabilistic structure. Second, we tested infants younger than those observed by Saffran et al., 1996, Saffran et al., 1999, to probe the developmental time-course of statistical learning during the first year after birth. We used a visual habituation procedure, an effective tool for investigating perceptual and cognitive processes in infants as young as neonates (Slater, 1995).