Pattern Learning in Infants and Neural Networks

Two recent sets of experiments have greatly expanded our understanding of the capacity of infants to detect regularities in patterns. Saffran, Aslin, and Newport (1996) conducted a set of experiments to investigate the sensitivity of infants to the statistical properties of patterns. Eight-month-olds heard strings of syllables consisting of randomly concatenated three-syllable “words”, sequences that never varied internally. Thus the transition probabilities within words were higher than between words. There were four different words consisting of twelve different syllables, and the subjects heard 180 words in all. Later the infants indicated, through differences in looking times, that they differentiated between these words and non-word three-syllable sequences which they had either heard with less frequency than the words (because they consisted of sequences of syllables which crossed word boundaries) or not heard at all. This is taken as evidence that they had picked up the statistics in the training set. More recent experiments (Saffran et al., 1999) have achieved similar results with patterns consisting of tones of different pitches. Marcus, Vijayan, Bandi Rao, and Vishton (1999) conducted a set of experiments to determine whether infants could extract rules from a pattern. Seven-month-olds were presented with series of three-syllable sequences separated by gaps. Each sequence consisted of two different syllables arranged in a fixed pattern, (x x y), (x y y), or (x y x). For example, in the (x y y) condition, the presented patterns included sequences such as le di di and ji je je. During the familiarization phase, each of 16 sequences was presented three times. Later the infants were tested on sequences consisting of novel syllables, in one experiment, syllables designed so as not to overlap with the training syllables. The test sequences either obeyed the training rule, or they obeyed one of the other two rules. The infants indicated, through differences in looking times, that they differentiated the patterns that followed the training rule from those that did not. This is taken as evidence that they had in some sense picked up the rule implicit in the training patterns.The experiments of Marcus et al. have attracted a good deal of attention because the authors, as well as Pinker (1999), have seen them as a challenge to connectionist models. Because connectionist models can only respond on the basis of similarity to training items, they apparently could not generalize to test sequences consisting of novel syllables, as the infants did. We believe that the line is not a simple one separating symbolic and connectionist models. The question is: what minimal set of mechanisms is required to achieve a behaviour, and what sorts of predictions a model embodying these mechanisms makes. In this paper we present an account of patterns and pattern learning that brings together the two sorts of experiments in a single framework, and we propose a minimal set of mechanisms based on this account that could achieve the sort of pattern learning we see in the experiments. We also show how a connectionist implementation of these mechanisms models the results of the experiments and makes novel predictions. This is not the first connectionist model of the Marcus et al. results (see, for example, Christiansen, Conway, & Curtin, 2000; Seidenberg et al., 1999; Shultz & Bale, 2000); however, we believe that the model we propose has the advantage of simplicity and perspicuity.