What Makes a Good Example? Modeling Exemplar Selection with Neural Network Representations

TL;DR

This work investigates how humans select informative exemplars for teaching, examining trade-offs between representativeness and diversity. It grounds the analysis in neural-network feature spaces (ResNet-50 and ViT-B/16) and formal subset-selection objectives (prototypicality, representativity, diversity, and their combinations) to predict exemplar choices. Across three novel 1D morph categories, data show that humans optimize a joint representativity–diversity objective, with transformer representations aligning more closely to human judgments than CNNs. The findings connect dataset-distillation ideas to cognitive theories of teaching, suggesting practical implications for computational models that distill informative exemplars and for understanding how learners structure information to facilitate learning.

Abstract

Teaching requires distilling a rich category distribution into a small set of informative exemplars. Although prior work shows that humans consider both representativeness and diversity when teaching, the computational principles underlying these tradeoffs remain unclear. We address this gap by modeling human exemplar selection using neural network feature representations and principled subset selection criteria. Novel visual categories were embedded along a one-dimensional morph continuum using pretrained vision models, and selection strategies varied in their emphasis on prototypicality, joint representativeness, and diversity. Adult participants selected one to three exemplars to teach a learner. Model-human comparisons revealed that strategies based on joint representativeness, or its combination with diversity, best captured human judgments, whereas purely prototypical or diversity-based strategies performed worse. Moreover, transformer-based representations consistently aligned more closely with human behavior than convolutional networks. These results highlight the potential utility of dataset distillation methods in machine learning as computational models for teaching.

Figures (4)

• Figure 1: For each stimulus category (dax, vep, bem), we compute the cosine similarity between the ResNet feature representation of each member and the representation of the midpoint stimulus (dashed vertical line). We also visualize the left and right endpoints along with the midpoint.
• Figure 2: Prototypicality of selected exemplars by condition. Dotted horizontal line indicates chance performance. As the quota for the number of exemplars increases, the prototypicality score tends to increase, meaning that the chosen exemplars become less typical and examples at the endpoints are chosen more frequently.
• Figure 3: ViT exemplar selection by choice criterion (blue lines) overlaid on human selections.
• Figure 4: The mean absolute error of the prototypicality score and diversity score between the neural network model and human participants, collapsed across all participants and exemplar quota condition. For the diversity score, the representativity criterion predicts human behavior with an error of $0.071$ and $0.072$ for ViT and ResNet, respectively. For the prototypicality score, the combined criterion yields an error of $0.066$ and $0.095$ for ViT and ResNet, respectively.