Patterson & Karuza (2020) Schrödinger's category: Active learning in the face of label ambiguity

Abstract

Research on active category learning–i.e., where the learner manipulates continuous feature dimensions of novel objects and receives labels for their self-generated exemplars–has routinely shown that people prefer to sample from regions of the space with high class uncertainty (near category boundaries). Prevailing accounts suggest this strategy faclitates an understanding of the subtle distinctions between categories. However, prior work has focused on situations where category boundaries are rigid. In actuality, the boundaries between natural categories are often fuzzy or unclear. Here, we ask: do learners pursue uncertainty sampling when labels at the boundary are themselves uncertain? To answer this question, we introduce a fuzzy buffer around a target category where conflicting labels are returned from two teachers, then we evaluate how sampling and representation are affected. We find that, relative to the rigid boundary case, learners avoid uncertainty, opting to sample densely from highly certain regions of the target category as opposed to its boundary. Subsequent generalization tests reveal that the sampling strategies encouraged by the fuzzy boundary negatively affected participants’ grasp of the category structure, even outside the fuzzy buffer zone.