4-month-olds Discover Algebraic Patterns in Music That 7.5-month-olds Do Not

4-month-olds Discover Algebraic Patterns in Music That 7.5-month-olds Do Not Colin Dawson (CDawson@Email.Arizona.Edu) Department of Psychology, 1503 E. University Blvd Tucson, AZ 85719 LouAnn Gerken, Ph.D. (Gerken@U.Arizona.Edu) Department of Psychology, 1503 E. University Blvd Tucson, AZ 85719 possesses a hierarchical structure, with units of different sizes nested within each other. Whereas in language the set of nested units might include phonemes, syllables and words, in music the units are notes, chords and phrases of various lengths. The input can contain structure at each of these levels; which level is the most salient will depend in part on the expectations of the listener. The second question addressed by this paper is whether, over time, a learner’s expectations will come to correspond to the sorts of structure frequently encountered in a particular domain. Abstract The recent language acquisition literature has revealed powerful, domain-general learning mechanisms that rely on frequency and distributional properties of the input. A second set of research investigates generalization of learned patterns to new input. The relationship between these two bodies is not entirely clear. The present paper demonstrates that the latter mechanism appears to be domain-general as well, although, in the domain of musical chord sequences, there is a decrease in sensitivity to algebraic patterns between 4 months of age and 7.5 months of age. We show that 4-month-old infants can discriminate three-chord patterns based on their AAB or ABA structure after being exposed to a set of patterns representing one of the structures. By 7.5 months, however, infants do not seem to make the discrimination, probably due to their experience with musical sequences, in which repetition plays less of a role than tonal structure. What is Statistical Learning? Introduction To begin with, a brief discussion of the general cognitive ability in question is warranted. The phenomenon referred to as “statistical learning” (SL) has been of great interest to those studying cognitive development over the past decade. Broadly construed, SL refers to an unsupervised learning process wherein an organism extracts frequency, probability and/or distributional information from a set of input, thereby structuring perception of the input. One set of SL studies concerns subjects’ ability to parse input into constituent elements – a critically important skill for learning language. Saffran, Newport and Aslin (1996) and Saffran, Aslin and Newport (1996) familiarized adults, pre-school-aged children and 8-month-old infants with a continuous stream of syllables comprising tokens of several back-to-back, three-syllable pseudowords drawn from a given “vocabulary”. When tested, all three groups could discriminate between constituent strings and those that occurred in the input stream, but which spanned word boundaries. In order to successfully perform this discrimination, subjects had to track some sort of frequency information in the familiarization stream. Further research suggested that the relevant information might be the set of transitional probabilities between elements (syllables and/or segments), and not simple pattern frequency (Aslin, Saffran, & Newport, 1998). Maye, Werker and Gerken (2002) have observed infants making use of a different kind of frequency information. They show that infants make use of distributional properties of the input, appearing to form either one or two phonetic categories depending on whether instances of a consonant in a familiarization stream have a unimodal or bimodal distribution with respect to voice onset time. A second set of research examines learners’ abilities to detect relational properties in the input, and to discriminate Until recently, it was thought in language acquisition circles that the linguistic input a child receives is much too impoverished to provide her with the rich structure necessary to support observed production; that is, unless her genetic endowment provided an intricate scaffolding in the form of a Universal Grammar. On this view, much of the structure was built in, and all that remained was for a set of “switches” to become “triggered” by a few relevant examples. It has now become clear that the linguistic input available to the child is vastly richer in information than previously believed. Much of this information exists as statistically reliable relationships among components of the input. This discovery has led to a burgeoning line of research in cognitive science, dedicated to examining a phenomenon labeled “statistical learning”. One of the questions raised by the research on statistical learning is whether the same learning mechanisms apply across a range of domains and not just to language. The first question directly addressed by the experiment discussed here is whether the ability to make a generalization based on the abstract relational properties contained in an input set is unique to the domain of language, or whether it might be applied to analogous structure in musical input. In order to discover statistical structure pertaining to abstract relational properties across the units in an input set, the learner must be inclined to analyze the input with respect to that particular unit. Music, like language,