Ignore Similarity If You Can: A Computational Exploration of Exemplar Similarity Effects on Rule Application

Abstract

It is generally assumed that when making categorization judgments the cognitive system learns to focus on stimuli features that are relevant for making an accurate judgment. This is a key feature of hybrid categorization systems, which selectively weight the use of exemplar-and rule-based processes. In contrast, Hahn et al. (2010) have shown that people cannot help but pay attention to exemplar similarity, even when doing so leads to classification errors. This paper tests, through a series of computer simulations, whether a hybrid categorization model developed in the ACT- showed that while the presence of an exemplar-similarity effect on response times was associated with classification errors it was possible to fit both measures to the observed data for an unsupervised version of the task (i.e., in which no feedback on accuracy was given). Difficulties arose when the model was applied to a supervised version of the task in which explicit feedback on accuracy was given. Modeling results show that the exemplar-similarity effect is diminished by feedback as the model learns to avoid the error-prone exemplar-route, taking instead the accurate rule-route. In contrast to the model, Hahn et al. found that people continue to exhibit robust exemplar-similarity effects even when given feedback. This work highlights a challenge for understanding how and why people combine rules and exemplars when making categorization decisions.

Building Memory Representations for Exemplar-Based Judgment: A Role for Ventral Precuneus

“…This result could imply that participants did not engage in initial hypothesis testing, but rather memorized exemplars immediately, supported also with the parametric increase in the ventral precuneus and the vmPFC. Similarity-based processes have been shown to be hard to resist (Brooks and Hannah, 2006; Hahn et al, 2010; von Helversen et al, 2014a; Brumby and Hahn, 2017; Bröder et al, 2017) and has been argued to be an unavoidable consequence of attending a stimulus in instance learning (Logan, 1988) which might suggest that an exemplar-based strategy was consistently used. Alternatively, the model fit measure used here is rather coarse, and fine-grained behavioral differences connected to a potential rule-bias within a learning block might have been missed.…”

“…Medin and Schaffer, 1978; Nosofsky, 1984; see e.g., Juslin et al, 2003, 2008). It has even been suggested that similarity-based strategies are hard to resist when making inferences (Brooks and Hannah, 2006; Hahn et al, 2010; von Helversen et al, 2014a; Brumby and Hahn, 2017; Bröder et al, 2017).…”

A cognitive category-learning model of rule abstraction, attention learning, and contextual modulation.

“…We assume that both systems are always active (see Brumby & Hahn, 2017;Hahn, Prat-Sala, Pothos, & Brumby, 2010;Lacroix, Giguere, & Larochelle, 2005; ; but see next Equation). Including normalization constants (among others) prevents these predictions to become 0 or 1.…”

A Cognitive Category-Learning Model of Rule Abstraction, Attention Learning, and Contextual Modulation

“…It is noteworthy that other hybrid models use a weighting or trade-off parameter that defines which module dominates the decision (e.g., ATRIUM Erickson & Kruschke, 1998; see further Pothos & Wills, 2011). In contrast, we assume that both systems are always active (see Brumby & Hahn, 2017;Hahn, Prat-Sala, Pothos, & Brumby, 2010), and that their input is additive. This, however, does not imply that both rules and configural memory will always equally contribute to the predictions, as this depends on what CAL learns.…”