Aslin, R.
A simplifying assumption in most experiments on statistical learning is that the structures in the
environment which are available for learning are consistently present in sub-samples of the input.
Given that input in the real world has considerable variability from sub-sample to sub-sample, the
naive learner is faced with the following dilemma. Either (1) extensively sample a representative
corpus, thereby ensuring that all structures, sub-structures, and contexts are incorporated into
the internal model of the environment (but running the risk of over-fitting), or (2) assume that there is only a single structure to be learned, thereby running the risk that sub-structures and contexts will be missed or that earlier sub-samples will result in garden-path effects. I will review several lines of research on statistical learning in a changing environment, including word segmentation from fluent speech, phonetic category learning, serial reaction-times to visual stimuli, and artificial grammar learning, that have been studied in human infants, human adults, and non-human animals.
I will draw comparisons with the reinforcement-learning literature, that makes a distinction between model-based and model-free learning, in an attempt to bridge to the statistical learning literature, including a discussion of the neural mechanisms that mediate performance in the paradigms used in these two literatures. A key feature of this comparison is how models of statistical learning, in the absence of overt reward, should conceptualize the role of attention, information seeking, and hypothesis generation.