Cobweb: An Incremental and Hierarchical Model of Human-Like Category Learning

TL;DR

Cobweb addresses how humans learn categories from limited exemplars by constructing incremental, hierarchical trees guided by Category Utility ($CU$), defined as $CU(c) = P(c)[U(c_p)-U(c)]$, where $U(c)$ is the feature-uncertainty at concept $c$ and $c_p$ is its parent. The paper demonstrates that Cobweb can flexibly span prototype-like and exemplar-like representations within a single framework and can incrementally learn across training blocks. By evaluating Cobweb against classic experiments (Medin1978context and Shepard1961learning) and by comparing leaf versus basic-level predictions, the study provides evidence of alignment with human categorization while revealing limitations and directions for future work, including naturalistic stimuli and image-based extensions. Overall, Cobweb emerges as a robust, hierarchical model of human category learning with potential to capture a broad range of learning phenomena.

Abstract

Cobweb, a human-like category learning system, differs from most cognitive science models in incrementally constructing hierarchically organized tree-like structures guided by the category utility measure. Prior studies have shown that Cobweb can capture psychological effects such as basic-level, typicality, and fan effects. However, a broader evaluation of Cobweb as a model of human categorization remains lacking. The current study addresses this gap. It establishes Cobweb's alignment with classical human category learning effects. It also explores Cobweb's flexibility to exhibit both exemplar- and prototype-like learning within a single framework. These findings set the stage for further research on Cobweb as a robust model of human category learning.

Figures (2)

• Figure 1: An illustrative example of Cobweb's learning process which involves incorporating a new instance, depicted as a dark green table, into the existing tree structure. Cobweb traverses the tree from the root to a specific leaf node, and along this path, the concept nodes (highlighted in light green) are updated to reflect the given instance. The changes resulting from fitting the new instance into the tree are denoted in bold and italics. During this traversal, Cobweb considers four available operations at each branch, adding, creating, merging, and splitting. It then proceeds with the operation that yields the highest averaged category utility within the subtree. For instance, for the branch in the red dot box, Cobweb chooses to add the instance to the "best" child because it results in the highest average utility score.
• Figure 2: The learning curves for human participants (red), the leaf prediction level (orange), and the basic prediction level (light green) across the learning blocks $1-23$. Learning block is on the $x$-axis and (human and model) accuracy on the $y$-axis.