Featural relations in concept learning and generalization

The feature-based concept learning literature has focused primarily on how subjects learn mappings from a set of stimulus features to a set of category labels. However, there are real-world concepts that cannot be predicted by feature values independently, but instead depend on knowledge or awareness of what we refer to as featural relations, i.e., labeled relationships that hold between the feature values within items. For example, in the domain of dating, the featural relation of an age gap would represent a difference greater than a certain amount between the ages of two people in a couple. Theoretical accounts of feature- and relation-based concept learning have remained largely independent due to the apparent gulf between the nature of flat and structured representations. There is little prior research that speaks to an intermediate and potentially bridging space such as feature-based representations that possess a limited structural aspect. In the present work, we identify featural relations as a promising middleground focusing on specific properties of relative magnitude that hold between feature values. We conducted three experiments that show: (1) the psychological validity of featural relations; (2) that theoretical accounts of feature-based categorization predict some aspects of human learning and generalization in the domain of concepts defined by featural relations – but only if it is assumed that between-feature comparisons are available as direct cues for learning; and (3) evidence of generalization behavior that is not predicted by feature-based theories but does correspond with relational cognition as well as prior findings in the reinforcement learning literature.