Constructing Deep Concepts through Shallow Search

Bonan Zhao, Christopher G. Lucas, Neil R. Bramley

Introduction

Humans manage to grasp complex concepts despite severe cognitive limitations by using "bootstrap learning"—a process of building, caching, and reusing simple sub-concepts to reach deeper insights. This paper argues that while this "shallow search" strategy is efficient, it makes human learning highly path-dependent. The order in which we encounter information determines whether we achieve a breakthrough or fall into a "learning trap."

The Resource-Rational Library Learning Framework

Modern AI often scales by increasing model size and compute, but human cognition operates under strict "resource-rational" constraints. The authors propose a framework that combines symbolic programs with a cache-and-reuse mechanism inspired by Bayesian library learning.

Bootstrap learning under limited computation

The core innovation is the introduction of a shallow search depth cap. Unlike models that search an infinite space, this model mimics human limitations by searching only locally. To solve complex problems, the model must first find and store simple components in its library, which then "shortens" the search distance for more complex concepts that build upon them.

Experimental Design and Datasets

The paper relies on behavioral data from experiments where participants observed animated evidence generated by a hidden ground-truth rule: $R' \leftarrow stripe(A) \times R - spot(A)$.

The Construct vs. De-construct experimental conditions

The conditions differed only in the order of evidence:

  • The Construct Group: Saw simple evidence (stripes) before complex evidence (spots).
  • The De-construct Group: Saw the complex evidence first.

Key Results and Insights

The results highlight that human learning is not commutative; seeing A then B is not the same as seeing B then A.

  • Path Dependency: Participants in the "Construct" condition were significantly more likely to discover the ground truth. Those in the "De-construct" condition struggled, often inducing convoluted, "messy" rules.
  • Learning Traps: Bootstrapping is a "free lunch" that allows us to reach beyond our cognitive grasp, but it creates "learning traps." Once an inappropriate sub-concept is cached, it is hard to "unlearn" and can actively impair future learning.
  • Shallow Search, Deep Results: You don't need a deep search algorithm if you have a good library. By reusing previously learned modules, a shallow search can produce incredibly sophisticated programs.
  • Model Fit: The rational library learning model provided a better quantitative fit to human generalization data than "naive" Bayesian models that lack these constraints.

Discussion and Future Work

This research provides computational evidence for why "facilitatory curricula"—starting with simple examples—are essential for human-like inductive inference. Future work could explore how humans prune or restructure mental libraries when foundational concepts are wrong, and how to scale these symbolic approaches to messy, high-dimensional real-world data. Identifying where conceptual paths diverge is also key to building AI teaching assistants that can align with a student's specific mental state.

Summary updated

Abstract

We propose bootstrap learning as a computational account for why human learning is modular and incremental, and identify key components of bootstrap learning that allow artificial systems to learn more like people. Originated from developmental psychology, bootstrap learning refers to people's ability to extend and repurpose existing knowledge to create new and more powerful ideas. We view bootstrap learning as a solution of how cognitively-bounded reasoners grasp complex environmental dynamics that are far beyond their initial capacity, by searching ‘locally’ and recursively to extend their existing knowledge. Drawing from techniques of Bayesian library learning and resource rational analysis, we propose a computational modeling framework that achieves human-like bootstrap learning performance in inductive conceptual inference. In addition, we demonstrate modeling and behavioral evidence that highlights the double-edged sword of bootstrap learning, such that people processing the same information in different batch orders could induce drastically different causal conclusions and generalizations, as a result of the different sub-concepts they construct in earlier stages of learning.
BibTeX
@inproceedings{zhao2024constructing,
  author = {Bonan Zhao and Christopher G. Lucas and Neil R. Bramley},
  title = {{Constructing Deep Concepts through Shallow Search}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaaiss.v3i1.31292},
}
Figure