Models of Human Category Learning: Do They Generalize?

Abstract

Models of Human Category Learning: Do They Generalize? Nolan Conaway (nconawa1@binghamton.edu) Kenneth J. Kurtz (kkurtz@binghamton.edu) Department of Psychology, Binghamton University Binghamton, NY 13905 USA Abstract Generalization to new examples is an essential aspect of categorization. However, recent category learning research has not focused on how people generalize their category knowledge. Taking generalization to be a critical basis for evaluating formal models of category learning, we employed a ‘minimal case’ approach to begin a systematic investigation of generalization. Human participants received supervised training on a two-way artificial classification task based on two dimensions that were each perfect predictors. Learners were then asked to classify new examples sampled from the stimulus space. Most participants based their judgments on one or the other dimension. Varying the relative levels of dimension salience influenced generalization outcomes, but varying category size (2, 4, or 8 items) did not. We fit two theoretically distinct similarity-based models (ALCOVE and DIVA) to aggregate learning data and tested on the generalization set. Both models could explain important aspects of human performance, but DIVA produced a superior overall account. Keywords: generalization; categorization; formal models of category learning; similarity; cognitive modeling. Introduction Categorization is an essential cognitive function – categories serve to organize knowledge and, critically, as a basis for extending knowledge to make sense of new experience. A full understanding of human categorization depends on developing models and theories that account for systematic patterns of human learning and generalization performance (for an overview of generalization, see Levering & Kurtz, 2010). In classic research, Roger Shepard (1957, 1987) put forth the idea of a universal law in which stimulus generalization follows an exponential function of distance in psychological space. This work has had broad implications for theoretical models of categorization. Highly influential reference point models (such as the exemplar view) compute classification in a manner that closely follows Shepard’s proposal. Specifically, the class membership of a known item is likely to be generalized to a new item if the two items are highly similar. The key additional design feature needed to account for human classification performance is the inclusion of a selective attention mechanism such that particular dimensions can matter more or less in the computation of similarity. Generalization performance (classification of previously unseen items) has been one of the most important important testing grounds in the debate between exemplar- and prototype-based accounts (e.g., Homa, 1984; Nosofsky, 1992; see also Medin & Schaffer, 1978 and the ensuing literature on behavioral experimentation and model-fitting with the 5-4 classification problem). In a somewhat different approach to studying the generalization of category knowledge, researchers have investigated whether exemplar models can account for rule- like generalization after category learning (Erikson & Kruschke, 1998, 2002; Nosofsky & Johansen, 2000). In these studies, participants were asked to classify novel instances after learning an artificial two-way classification based on a unidimensional rule with exceptions. The critical test items were highly similar to the exceptions, but clearly classifiable using the rule. The outcomes of these studies were somewhat mixed and appear to depend on stimulus attributes and also on the structure of the categories that are learned. The goal of the present research is two-fold: 1) to explore a different approach to investigating the psychology of category generalization; and 2) to use generalization performance as a basis to compare and differentiate models that are highly successful in fitting human learning data. Toward the first goal, our experimental approach is broadly comparable to the psychological studies of generalization discussed above: after a learning phase, participants are asked to classify novel examples. However, our work differs in that we use minimal category learning conditions (small numbers of examples that are readily assigned to two fully coherent classes). Our primary aim is to identify basic, systematic properties of generalization performance. Regarding the second goal, the field presently offers a small group of formal models of category learning that are general purpose (applicable to any classification problem), that provide explanation at the level of process/mechanism, and that yield good fits to established benchmarks for human category learning. Within the realm of fitting human classification learning performance, there is some sense of having hit the ceiling in terms of differentiating among these models despite their having distinct explanatory elements. Our rationale is that models that do quite well in fitting learning data may diverge in their ability to account for patterns of generalization performance. In particular we are compelled by the prospect of fitting model parameters to the learning data and then holding the models to these values in evaluating ensuing generalization (as discussed below). Toward this end, we evaluate two successful models: a canonical representative of the reference point approach, ALCOVE (Kruschke, 1992) and an updated