Tom Griffiths: UC Berkeley Department of Psychology and Group Major in Cognitive Science

Learning a language, a concept, or a social norm all require making an inductive inference, going beyond the data provided by the behavior of other people. Such inferences are underdetermined by the available data, allowing people’s biases to influence the conclusion that they reach. These inductive biases reflect the expectations that people have about which hypotheses are most likely to account for the data they observe, and the factors that make it easier to learn some languages, concepts, and norms rather than others. I will present a combination of theoretical results and laboratory experiments demonstrating that inductive biases can have a significant effect on the changes that information undergoes when it is passed from one person to another. By simplifying the process of cultural transmission down to a “diffusion chain” along which information is passed, we can examine how the inductive biases of individual agents influence the outcome of this transmission process. Assuming that the agents passing information are rational Bayesian learners, we can prove that over time the information they transmit will be transformed into a form more consistent with their inductive biases. Laboratory experiments in which diffusion chains are constructed from human participants bear out this prediction, with experiments in which a sequence of people learn from data generated by other people or reconstruct information from memory producing results that reflect their inductive biases. I will discuss the implications of these results for understanding cultural evolution more generally, as well as for identifying the circumstances under which we can expect knowledge to accumulate over time.
http://cocosci.berkeley.edu/tom/papers/ilreview.pdf