Comparing methods of category learning: Classification versus feature inference

Abstract

AbstractCategories have at least two main functions: classification of instances and feature inference. Classification involves assigning an instance to a category, and feature inference involves predicting a feature for a category instance. Correspondingly, categories can be learned in two distinct ways, by classification and feature inference. A typical difference between these in the perceptual category learning paradigm is the presence of the category label as part of the stimulus in feature inference learning and not in classification learning. So we hypothesized a label-induced rule-bias in feature inference learning compared to classification and evaluated it on an important starting point in the field for category learning – the category structures from Shepard, Hovland, and Jenkins (Psychological Monographs: General and Applied, 75(13), 1-42, 1961). They classically found that classification learning of structures consistent with more complex rules resulted in poorer learning. We compared feature inference learning of these structures with classification learning and found differences between the learning tasks supporting the label-bias hypothesis in terms of an emphasis on label-based rules in feature inference. Importantly, participants’ self-reported rules were largely consistent with their task performance and indicated the preponderance of rule representation in both tasks. So, while the results do not support a difference in the kind of representation for the two learning tasks, the presence of category labels in feature inference tended to focus rule formation. The results also highlight the specialized nature of the classic Shepard et al. (1961) stimuli in terms of being especially conducive to the formation of compact verbal rules.