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

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.

💡 Research Summary

This paper investigates how people choose a small set of exemplars to teach a novel visual category, focusing on the trade‑off between representativeness (or prototypicality) and diversity. The authors bridge cognitive‑science questions about pedagogical example selection with machine‑learning techniques from dataset distillation and subset selection. They first construct three artificial categories—“dax”, “vep”, and “bem”—each consisting of a one‑dimensional morph continuum ranging from 0 to 100. Images at intermediate values are visually prototypical, while those at the extremes are highly diverse.

Feature representations for every image are extracted from two pretrained vision models: a convolutional ResNet‑50 and a transformer‑based ViT‑B/16, both trained on ImageNet. After L2‑normalisation, cosine similarity defines a distance matrix in each model’s embedding space. Using these distances, the authors formalise four exemplar‑selection objectives:

1. Prototypicality – rank individual items by their average similarity to the whole set (or to the midpoint) and pick the top‑M items.
2. Representativeness – a facility‑location formulation that maximises the sum over all data points of the similarity to their nearest selected exemplar, encouraging coverage while penalising redundancy.
3. Diversity – maximise the sum of pairwise distances among the selected items, pushing the set toward the extremes of the distribution.
4. Combined (Representativeness + Diversity) – a simple equal‑weight sum of the two previous objectives, intended to capture a balanced teaching strategy.

Twenty‑four adult participants were recruited via a university pool. In each trial they imagined being an astrobiologist tasked with teaching trainees about an alien species. For each of the three categories they were allowed to select either one, two, or three exemplars (quota varied across blocks). Their selections were recorded and analysed using two behavioural metrics: a prototypicality score (average absolute distance of chosen items from the midpoint, normalised) and a diversity score (maximum pairwise distance among chosen items, normalised). Chance‑level expectations were computed analytically for both metrics.

Key behavioural findings: when only one exemplar could be chosen, participants tended to pick the midpoint, performing at chance level for prototypicality. With two or three exemplars, prototypicality scores rose significantly, indicating a shift toward including more peripheral items. Diversity scores were also substantially above chance (22.5 % above for two exemplars, 33.8 % above for three), showing a clear preference for spreading selections across the category range.

Model‑human comparisons revealed that the pure prototypicality objective performed worst, failing to capture the observed increase in diversity with larger quotas. The representativeness objective aligned most closely with human choices, especially when using the ViT‑B/16 embeddings, suggesting that global self‑attention representations mirror human judgments about what constitutes a “good teaching set”. The combined objective also matched human behaviour reasonably well, though the simple equal‑weight sum sometimes over‑emphasised diversity relative to the human data.

Overall, the study concludes that people intuitively balance coverage of the whole distribution with avoidance of redundancy, and they become more diversity‑oriented when allowed to present multiple examples. Moreover, transformer‑based visual features provide a better computational proxy for human exemplar selection than traditional convolutional features. The authors argue that these findings link dataset distillation research with cognitive theories of teaching, offering a promising avenue for designing machine‑learning systems that emulate human pedagogical strategies.