Does generalization in infant learning implicate abstract algebra-like rules?

Two recent papers in Science [1, 2] suggest that infants well under a year old can learn from exposure to relatively short samples of language-like sequences of syllables. The first of these two papers, by Saffran and colleagues [1], suggested that infants were sensitive to the statistical structure of syllable sequences that they heard, in that they tended to listen longer to syllable sequences that were less common in their brief experience. An interesting exchange of commentaries ensued (Science, 276, 1177-1181), centering mostly on the issue of whether it is surprising that infants are sensitive to the statistical structure of experience. No one has admitted to being surprised; what the debate has been about is whether there are any theoretical positions under which anyone should have been surprised. We think enough has been said about that. The findings reported in the second of these papers, by Marcus and colleagues [2], have raised a rather different debate, this one centering on whether or not infants learn something over and above the statistical structure of the language-like sequences that they hear. Specifically, the authors found that, following exposure to syllable sequences obeying a particular pattern, seven-month-old infants tended to listen longer to syllable sequences which violated the pattern, even when the new sequences were composed of novel syllables. The authors suggest that their findings cannot be accounted for by any learning method that relies on statistical information. Rather, they claim, the results implicate the learning and use of abstract ”algebra-like” rules. Before we review the findings in more detail, we will lay our cards on the table: We don’t think the findings of Marcus and colleagues really cut any ice at all regarding the possible existence of abstract rule learning in infants. We’ll support this position by describing several ways in which the data might be seen as compatible with the extraction of statistical information rather than abstract rules. We do think the question of the extent of infants’ generalization ability is very interesting, because it tells us something about what they bring to learning situations at a relatively early age. Indeed, it seems very possible to us that seven-month-old infants possess mechanisms that provide powerful support for generalization. But we don’t really see how experiments of this general sort can tell us whether they use rules per se; the powerful mechanisms might simply be ones that help statistical learning procedures generalize in powerful ways. Furthermore, these mechanisms might themselves be learned. Let us explain.