Feature, Alignment, and Supervision in Category Learning: A Comparative Approach with Children and Neural Networks

Understanding how humans and machines learn from sparse data is central to cognitive science and machine learning. Using a species-fair design, we compare children and convolutional neural networks (CNNs) in a few-shot semi-supervised category learning task. Both learners are exposed to novel object categories under identical conditions. Learners receive mixtures of labeled and unlabeled exemplars while we vary supervision (1/3/6 labels), target feature (size, shape, pattern), and perceptual alignment (high/low). We find that children generalize rapidly from minimal labels but show strong feature-specific biases and sensitivity to alignment. CNNs show a different interaction profile: added supervision improves performance, but both alignment and feature structure moderate the impact additional supervision has on learning. These results show that human-model comparisons must be drawn under the right conditions, emphasizing interactions among supervision, feature structure, and alignment rather than overall accuracy.

💡 Research Summary

This paper presents a rigorous comparative study investigating how children and convolutional neural networks (CNNs) learn object categories from limited data. Employing a “species-fair” design, the researchers exposed both learners to identical conditions in a few-shot, semi-supervised learning paradigm. The study systematically manipulated three key factors: the level of supervision (1, 3, or 6 labeled exemplars out of 6 total), the target feature defining category membership (size, shape, or pattern), and the perceptual alignment between exemplars (high or low). Alignment was operationalized based on Structural Mapping Theory; high-alignment pairs differed only along the target feature dimension, facilitating comparison, while low-alignment pairs varied in additional irrelevant features, making the invariant category-defining rule harder to discern.

Human participants were 24 children aged 5-7 years. They underwent a learning phase where they viewed pairs of novel objects, with a subset of pairs receiving explicit labels (e.g., “These are both modis”), followed by a test phase requiring classification of new instances. The CNN model was a Siamese network with a shared ResNet-18 encoder, trained with a combined loss function: binary cross-entropy for supervised examples and a contrastive loss for both supervised and unsupervised pairs to encourage meaningful representations, closely mirroring the pairwise structure of the child’s task.

The results revealed fundamentally different learning profiles. Children demonstrated remarkable data efficiency, showing no significant main effect of the number of supervised examples. Their performance was high even with just one labeled exemplar, underscoring strong inductive biases. However, their learning was heavily influenced by feature type (shape > pattern > size) and alignment, performing significantly better on high-alignment trials.

In stark contrast, CNNs exhibited a more classical machine learning pattern: performance generally improved with more labeled data. Crucially, however, both alignment and feature type acted as moderators, influencing the degree to which additional supervision benefited learning. For instance, in some conditions (e.g., low-alignment, size), CNNs performed near-perfectly regardless of supervision level, while in others (e.g., high-alignment, shape), the gains from more labels were more modest. This suggests that the “difficulty” of a learning problem can be perceived oppositely by humans and CNNs.

The study concludes that direct comparisons between human and machine learning must move beyond aggregate accuracy scores. The findings emphasize the critical interaction between supervision, feature structure, and perceptual alignment. Human efficiency arises from powerful pre-existing biases and structural alignment sensitivity, enabling learning from minimal instruction. CNN performance is more directly tied to the quantity of explicit feedback and the statistical regularities in the data. Therefore, understanding human-like learning and developing robust AI requires a nuanced analysis of how these fundamental factors interact within each system, rather than seeking a single superior learner.