The transformational theory of similarity suggests that whenjudging similarity, people are sensitive to the number of trans-formation operations needed to make two compared repre-sentations match. Although this theory has been influential,little is known about how transformations are learned andto what extent learned transformations affect similarity judg-ments. This paper presents two experiments addressing thesequestions, in which people learned categories defined by atransformation. In Experiment 1, when the transformationswere directly visible, people had no trouble learning and ap-plied their knowledge to similarity and categorization judg-ments involving previously unseen items. In Experiment 2,the task required transformations to be inferred rather than ob-served. People were still able to learn the categories, but inthis more difficult case ratings were less strongly affected bytraining. Overall, this work suggests that newly learned trans-formations can impact similarity judgments but the salience ofthe transformation has a large impact on transfer.

Human activity requires an ability to generalise beyond theavailable evidence, but when examples are limited – as theynearly always are – the problem of how to do so becomes par-ticularly acute. In addressing this problem, Shepard (1987)established the importance of representation, and subsequentwork explored how representations shift as new data is ob-served. A different strand of work extending the Bayesianframework of Tenenbaum and Griffiths (2001) established theimportance of sampling assumptions in generalisation as well.Here we present evidence to suggest that these two issuesshould be considered jointly. We report two experiments whichreveal replicable qualitative patterns of individual differencesin the representation of a single category, while also showingthat sampling assumptions interact with these to drive gener-alisation. Our results demonstrate that how people shift theircategory representation depends upon their sampling assump-tions, and that these representational shifts drive much of theobserved learning.

What mechanisms underlie people’s ability to use cross-situational statistics to learn the meanings of words? Here wepresent a large-scale evaluation of two major models of cross-situational learning: associative (Kachergis, Yu, & Shiffrin,2012a) and hypothesis testing (Trueswell, Medina, Hafri, &Gleitman, 2013). We fit each model individually to over 1500participants across seven experiments with a wide range ofconditions. We find that the associative model better capturesthe full range of individual differences and conditions whenlearning is cross-situational, although the hypothesis testingapproach outperforms it when there is no referential ambiguityduring training.

Does the similarity between two items change over time? Previous studies (Goldstone & Medin, 1994; Gentner & Brem, 1999) have found suggestive results but have relied on interpreting complex interaction effects from “deadline” decision tasks in which the decision making process is not well understood (Luce, 1986). Using a self-paced simple decision task in which the similarity between two items can be isolated from strategic decision processes using computational modeling techniques (Ratcliff, 1978), we show strong evidence that the similarity between two items changes over time and shifts in systematic ways. The change in similarity from early to late processing in Experiment 1 is consistent with the theory of structural alignment (Gentner, 1983; Goldstone & Medin, 1994), and Experiment 2 demonstrates evidence for a stronger influence of thematic knowledge than taxonomic knowledge in early processing of word associations (Lin & Murphy, 2001).

How does the number of features impact category learning?One view suggests that additional features creates a “curse ofdimensionality” - where having more features causes the sizeof the search space to grow so quickly that discovering goodclassification rules becomes increasingly challenging. The op-posing view suggests that additional features provide a wealthof additional information which learners should be able to useto improve their classification performance. Previous researchexploring this issue appears to have produced conflicting re-sults: some find that learning improves with additional features(Hoffman & Murphy, 2006) while others find that it does not(Minda & Smith, 2001; Edgell et al., 1996). Here we inves-tigate the possibility that category structure may explain thisapparent discrepancy – that more features are useful in cate-gories with family resemblance structure, but are not (and mayeven be harmful) in more rule-based categories. We find whilethe impact of having many features does indeed depend on cat-egory structure, the results can be explained by a single unifiedmodel: one that attends to a single feature on any given trialand uses information learned from that particular feature tomake classification judgments.