The Role of Similarity in Generalization

摘要: The Role of Similarity in Generalization Matt Jones, W. Todd Maddox, and Bradley C. Love [mattj,maddox,love]@psy.utexas.edu University Texas, Department Psychology, 1 Station A8000 Austin, TX 78712 USA Abstract between stimulus differences category differences, making judgments. is often regarded as a fundamental construct underlying generalization learning many other domains. key assumption this approach that multidimensional stimuli are summarized by single value before entering the decision process. present study challenges showing judgments depend on full relationship past stimuli, way cannot be mediated unidimensional similarity measure. Approaches based response generalization, knowledge partitioning, distributional representations also shown to insufficient account for our findings. Recency Love, Maddox (2006) demonstrate how can directly measured during probabilistic classification task through analysis decisional recency effects. Specifically, they found responses biased towards feedback given previous trial, strength bias determined difference stimuli. effect thus represents from current one. (2005) when one dimension predictive label another irrelevant, becomes selectively dependent diagnostic (Fig. 1). This finding consistent with accounts selective attention assume adapts weight tas k- relevant dimensions more heavily (Kruschke, 1992; Nosofsky, 1986). In words, differing along become less similar than irrelevant dimension. adaptation observable effects, elaborated below. Introduction has long been held underlie wide range cognitive processes. Seminal work Shepard (1957) showed conditioning identification tasks explained terms since extended tasks, including categorization (Medin & Schaffer, 1978) inductive reasoning (Osherson, Smith, Wilkie, Lopez, Shafir, 1990). An important not constant; rather, it changes systematically function which attributes (Heit Rubinstein, 1994; However, still generally assumed well defined any judgment, context, attentional set. critical holds only spatial models similarity, but feature-set (Tversky, 1977) approaches internal relational structure (Markman Gentner, 1993). similarity. We describe an experiment using four-category subjects must attend two simultaneously, use these sources information different ways. principal multiple gradients simultaneously active aspects judgment. Thus performance (even trial trial). argue failing assumes relationships reduced passed That is, acts mediator or sufficient statistic. Instead, appears people particular alignment D ia gno stic d iff ere nce ce ren iffe t lev Irre Figure 1: Selective 2-category (Jones et al., 2005, Expt. 1, Condition F). Horizontal axes indicate successive Vertical axis shows (in log-odds) stimulus.