The Benefits of Non-rational Learning

Ansgar Endress

City University London


Dealing with the world requires making inferences over and above what we perceive. Examples include rule-learning during language acquisition, but also apparently simple processes such as grouping elements into objects. Over the years, different research traditions have linked such abilities to various types of mechanisms, from general associative or algebraic mechanisms to Bayesian inferences. An alternative approach to such general mechanisms relies on perceptual or memory primitives, basic psychological mechanisms that support the acquisition of certain rules. Just as other animals have a variety of specialized learning mechanisms, humans might also draw on specialized primitives to make important inferences. Here, I report results on two such primitives – a sensitivity to identity relations and a sensitivity to edges of sequences. I contrast the acquisition of rules relying on these primitives with rational, Bayesian approaches to rule learning, and show that rule-learning deviates substantially from the predictions of Bayesian approaches. Further, I show that, in the case of perceptual inferences, such primitives might lead to behavior that seems more adaptive than the behavior of a rational learner.