Rule Learning in Humans and Animals

In recent years, artificial language learning experiments have revealed a rich and complex picture of the abilities of different species and different human age groups to discover simple patterns in sequences. In one influential study, Aslin et al. (1998) show that human infants use transitional probabilities (TP's), and not just co-occurrence frequencies, between adjacent syllables in a monotonous stream of speech to segment it into word-like units. Peña et al. (2002) presented human adults with a sequence of syllables composed of concatenated triplets of the form AXC, where A and C consistently co-occur, while X may vary. Tested for recognition after 100 exposures to the sequence, their subjects show no preference for either rule-following unattested sequences (AYC, with Y unobserved in that position; henceforth 'rule-words') or rule-breaking attested sequences (XCA or CAX, henceforth 'part-words'). After 300 exposures, however, subjects prefer part-words (time effect), while with merely 20 exposures but with subliminal pauses added between triplets, subjects prefer rule-words (micropause effect). These results are often interpreted as evidence for two different processes: a statistical mechanism that tracks transitional probabilities, and a rule mechanism for structure detection. Endress and Bonnatti (2006) emphasize that the time effect runs contrary to the prediction of single mechanism models, and thus supports their More-than-One-Mechanism (MoM) hypothesis. Toro and Trobalón (2005) perform similar experiments with rats, and report a number of qualitative differences with the human results. In particular, although the rats learn to discriminate between stimuli on the basis of co-occurrence frequencies, T&T report that they find no TP-effect and no rule learning (and hence no time effect and no micropause effect). In our work, we investigate through modelling whether the presented empirical results really rule out a single mechanism account for the results on humans as well as rodents. We define a probabilistic model which uses the Simple Good-Turing method to quantify a subject's willingness to generalize as the amount of probability mass that is reserved for unobserved sequences. We further model the probability that a subject will retain a particular subsequence (Pret(s) = A length(s)) or recognize it and hence increase its subjective count (Prec(s) = (1 – B activation(s))D #types) .