Events

Broadly construed, statistical learning involves finding predictive patterns based on experiences of property distributions. Psychologists have developed many competing accounts of this kind of induction from instances. Characterizing the phenomena in terms of statistical learning provides a framework for comparing, and hopefully unifying, across alternatives. I will discuss two varieties of statistical learning especially relevant to research on cognitive development. The first concerns learning discriminative versus generative models. Sometimes people learn very narrow, special-purpose relations among properties (discriminative, such as p(x|y). Other times people learn more complete, general-purpose relations (generative, such as p(x,y). One hypothesis is that young children may be disposed to learn generative models. Some of children’s errors or limitations on learning tasks may stem from their trying to learn something more general than intended by the experimenter/teacher. A second variety of statistical learning distinguishes evidential from transductive inferences. Are experiences treated as a sample useful for drawing conclusions about a population (evidential), or are experiences treated as the population to be described (transductive)? This distinction provides a particular perspective on the “similarity-based” versus “theory-based” debate. Similarity-based accounts maintain that people make transductive inferences; theory-based accounts maintain that people make evidential inferences. It is this distinction that makes empirical studies of children’s sensitivity to sampling so critical in the theory vs. similarity debate.