Olejarczuk, P. & Kapatsinski, V.
University of Oregon
In Bayesian approaches to learning, the prior specifies the expectations/beliefs that learners bring with them to the learning task. Patterns that violate the learners? expectations are harder to learn. Much recent research in phonology has attempted to identify the prior beliefs that affect phonology learning (Moreton & Pater, 2012). In contrast, the mechanism for updating these beliefs has received almost no attention. However, Wilson (2006) has suggested that learners suffer from a kind of confirmation bias, in that evidence inconsistent with the prior is systematically discounted. In this paper, we present evidence for the opposite kind of bias: surprising stimuli - defined as having low p(D) - draw attention and are particularly likely to effect a change in belief. In contrast, unsurprising stimuli may fail to shift beliefs at all. We show that English speakers exposed to an extreme version of the English stress pattern (deterministic Latin Stress) fail to shift their beliefs about stress relative to a pre-test. In contrast, those exposed to an unfamiliar (and typologically unattested) stress pattern (Anti-Latin) do rapidly shift beliefs. By modeling the results with a Bayesian statistical model, we show that normative belief updating underestimates the role of surprise in driving statistical learning.