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

TL;DR

This study compares how children and CNNs learn categorical structure from sparse data under matched semi-supervised conditions, manipulating the diagnostic feature, exemplar alignment, and supervision level. A semi-supervised, contrastive framework with a Siamese CNN is used alongside a child-learning paradigm to assess parallel and divergent learning strategies. Results show that children rely heavily on perceptual alignment and robust shape biases, with limited gains from additional labels under optimal conditions, whereas CNNs benefit from supervision but exhibit complex interactions with feature type and alignment. The findings highlight that meaningful human–model comparisons depend on examining interaction patterns across factors, not merely overall accuracy, informing how inductive biases are acquired and applied in artificial systems.

Abstract

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.

Figures (4)

• Figure 1: Examples of stimuli pairs used for each feature category (shape, size, pattern) and alignment (high, low).
• Figure 2: A pair of images is processed independently by a shared CNN $f_{\theta}$, producing embeddings for each image. Depending on whether the given pair of images have explicit supervision, these embeddings are used differently. In the supervised setting (dark blue line), a classification head $g_{\eta}$ maps each embedding to a logit, trained using a supervised loss on $(\theta,\eta)$ when labels are available. In the unsupervised setting (light blue line), the same embeddings are passed through a projection MLP $h_{\phi}$, and a contrastive loss on $(\theta,\phi)$ is used to encourage embeddings from same-class pairs to be similar and those from different-class pairs to be dissimilar. A contrastive loss is also used for supervised pairs to further improve the representations learned by the neural network.
• Figure 3: Predicted proportion correct as a function of the number of supervised trials, feature type, and alignment condition for child (left) and CNN (right), with error bars for 95% CI. Lines show model predictions and points indicate condition means. Dotted line indicates chance performance.
• Figure 4: Predicted performance of the CNN separated by feature, alignment, and level of supervision, with error bars for 95% CI. Lines show model predictions and points indicate condition means.