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

Technical Summary: Feature, Alignment, and Supervision in Category Learning

Problem Statement
A central challenge in cognitive science and machine learning is understanding how humans and machines learn from sparse data. While humans, particularly infants, demonstrate remarkable efficiency in one-shot or few-shot learning, machine learning algorithms typically require orders of magnitude more data. Although deep neural networks (DNNs) have achieved success in supervised and self-supervised regimes, it remains unclear whether their learning strategies and inductive biases align with human cognition. Existing literature often relies on cross-study comparisons or loosely matched tasks, making it difficult to isolate the specific sources of success or failure in each system. This study addresses the need for "species-fair" comparisons—where human-like constraints are placed on machines and machine-like constraints on humans—to determine how supervision, feature structure, and perceptual alignment interact to support category learning in children versus Convolutional Neural Networks (CNNs).

Methodology
The study employs a controlled, species-fair semi-supervised learning framework comparing 5- to 7-year-old children ($N=24$) and a Siamese ResNet-18 architecture.

• Task Design: Both learners were exposed to novel object categories defined by three distinct feature dimensions: size, shape, or pattern.
• Variables: The study systematically manipulated three factors:
  1. Supervision Level: The number of labeled exemplars provided (1/6, 3/6, or 6/6).
  2. Perceptual Alignment: High vs. Low alignment. In high-alignment conditions, paired exemplars shared all features except the target dimension, facilitating clear structural correspondence. In low-alignment conditions, exemplars varied in multiple dimensions, obscuring the invariant target feature.
  3. Feature Structure: The category-defining dimension (size, shape, pattern).
• Human Procedure: Children viewed pairs of stimuli (either "same" or "different" pairs) with varying levels of verbal supervision (e.g., "These are both [Category]" vs. "Only this is a [Category]"). They were subsequently tested on their ability to classify new stimuli.
• Machine Procedure: A Siamese network with a shared ResNet-18 encoder was trained using a multi-objective loss function. It combined binary cross-entropy for supervised trials (predicting object vs. non-object) with a contrastive loss for both supervised and unsupervised pairs (pushing same-class embeddings together and different-class embeddings apart). The network was trained for 200 epochs across 20 random seeds to match the counterbalancing of the human study.

Key Results
The study revealed divergent learning profiles between children and CNNs, characterized by distinct interactions among supervision, alignment, and feature type:

1. Sensitivity to Perceptual Alignment:

   • Children: Performance was highly sensitive to alignment. Children were significantly more accurate in high-alignment conditions. Crucially, alignment interacted with feature type: low alignment severely impaired learning for size-based and pattern-based categories but had a negligible effect on shape-based categories, reflecting a robust "shape bias."
   • CNNs: The main effect of alignment was weak and not statistically significant on its own. However, alignment interacted with supervision and feature type. For shape-based categories, high alignment allowed supervision to improve performance significantly, whereas low alignment rendered additional supervision ineffective. Conversely, CNNs were largely insensitive to alignment for size and pattern categories.

2. Feature Structure Biases:

   • Children: Showed a strong preference for shape, outperforming size and pattern. Size was the most difficult feature, particularly under low-alignment conditions where performance approached chance levels.
   • CNNs: Exhibited the inverse pattern. Shape was the most difficult feature for the network, while size and pattern were learned with high accuracy, often reaching ceiling performance even with minimal supervision. This aligns with prior findings that CNNs favor texture-based cues over abstract shape representations.

3. Role of Supervision:

   • Children: Showed diminishing returns from additional labeled examples once optimal conditions (high alignment, shape feature) were met. Their learning was driven more by the quality of the perceptual structure than the quantity of labels.
   • CNNs: Performance improved monotonically with increased supervision across all conditions. The benefit of additional labels was most pronounced for shape-based categories under high alignment.

4. Interaction Effects:

   • The study found that the impact of supervision on learning was not uniform but was gated by the interaction of feature type and alignment. For CNNs, supervision was less effective for learning shape categories when exemplars lacked perceptual coherence (low alignment).

Significance and Claims
The authors claim that this work provides the first controlled comparison of child and CNN learning in a semi-supervised regime, moving beyond overall accuracy metrics to analyze the mechanisms of learning.

• Species-Fair Comparison: The results demonstrate that human-model comparisons must be drawn under specific, principled conditions. Apparent similarities or differences in performance often hinge on the interaction between supervision, target features, and perceptual alignment rather than aggregate accuracy.
• Inductive Biases: While both systems can acquire biases (e.g., children's shape bias, CNNs' texture/size bias), the way these biases interact with other learning constraints differs fundamentally. Children rely on perceptual alignment to extract relational structure, whereas CNNs rely on local feature statistics that remain informative even when global spatial correspondences are disrupted.
• Evaluation of Models: The paper argues that evaluating computational models solely on whether they achieve human-level accuracy is misleading. A more informative approach is to assess whether a model reproduces the same pattern of tradeoffs and interaction effects as humans across principled manipulations. This reveals whether two learners succeed for the same reasons or different reasons, offering a deeper understanding of the inductive biases driving efficient generalization in both biological and artificial systems.