Active inference in concept learning

摘要: Active inference in concept learning Jonathan D. Nelson (jnelson@cogsci.ucsd.edu)* Joshua B. Tenenbaum (jbt@psych.stanford.edu)^ Javier R. Movellan (movellan@cogsci.ucsd.edu)* *Cognitive Science Department, UCSD La Jolla, CA 92093-0515 ^Psychology Stanford University Stanford, 94305 Abstract People are active experimenters, constantly seeking new information relevant to their goals. A reasonable approach gathering is ask questions and conduct experiments that minimize the expected state of uncertainty, or maximize gain, given current beliefs (Fedorov, 1972; MacKay, 1992; Oaksford & Chater, 1994). In this paper we present results on an exploratory experiment designed study people’s behavior a task. The suggest subjects’ may be explained well from point view Bayesian maximization. Introduction scientific inquiry everyday life, people seek out perceptual cognitive tasks. Whether performing uncover causal relationships, saccading informative areas visual scenes, turning towards surprising sound, actively relative Consider person foreign language, who notices particular word, “tikos,” used refer baby moose, penguin, cheetah. Based those examples, she attempt discover what tikos really means. Logically, there infinite number possibilities. For instance, could mean animals, simply even animals antique telephones. Yet few examples often enough for human learners form strong intuitions about meanings most likely. Suppose learner duck, adult telephone, inquire whether object “tikos.” What question would ask? Why do think pointing telephone not good idea, though logical view, phone very tikos? normative theoretical framework, try predict tasks model presented here, evaluate terms value. Formally, defined with respect probability model. Here use framework sense internal as distributions. order quantify value (in bits) person’s questions, first need her beliefs, way updated obtained. (1999, 2000) provides such While 2000); last authors paper, pilot study, found his described well, were some deviations between predictions beliefs. which describe below, based Tenenbaum’s original model, but extended ways reduce previously observed participants’ We formalize situation by using standard probabilistic notation: random variables represented capital letters, specific values taken small letters. variable C represents correct hidden trial. This directly observable participants; rather, they infer it basis example numbers consistent true concept. Notation “C=c” shorthand event takes c, e.g. (or “hypothesis”) prime numbers. represent subjects vector X. subject’s concepts probable prior presentation any function P(C=c). belief concept’s probability, after sees