Reproduce “A Rational Model of Dimension-reduced Human Categorization” on OpenResearch

About this paper

Yifan Hong, Chen Wang

This research introduces mPPCA (mixture of probabilistic principal component analyzers), a rational model of human categorization that incorporates dimension-reduced representations within a nonparametric Bayesian framework. The model effectively predicts human categorization and few-shot generalization behavior in high-dimensional and noisy environments, outperforming traditional exemplar and prototype models by better capturing graded human judgments and exhibiting higher accuracy on behavioral tasks.

Key insights

• Dimension reduction is theoretically beneficial for categorization when the contribution of a principal component to within-category variation is relatively small compared to its contribution to differentiating category prototypes.
• A hierarchical prior on feature dimensions allows categories to reuse and inherit common patterns of variation, enabling robust few-shot learning for both subcategories and new categories.
• Humans' ability to generalize from limited examples and focus on salient features can be explained by a rational model that intrinsically reduces dimensionality and shares generalization biases.

Summary by alphaXiv.