Optimizing the human learnability of abstract network representations

William Qian, Christopher W. Lynn, Andrei A. Klishin, Jennifer Stiso, Nicolas Christianson, Dani S. BassettProceedings of the National Academy of Sciences 2022 · DOI · arXiv

Abstract

Precisely how humans process relational patterns of information in knowledge, language, music, and society is not well understood. Prior work in the field of statistical learning has demonstrated that humans process such information by building internal models of the underlying network structure. However, these mental maps are often inaccurate due to limitations in human information processing. The existence of such limitations raises clear questions: Given a target network that one wishes for a human to learn, what network should one present to the human? Should one simply present the target network as-is, or should one emphasize certain parts of the network to proactively mitigate expected errors in learning? To investigate these questions, we study the optimization of network learnability in a computational model of human learning. Evaluating an array of synthetic and real-world networks, we find that learnability is enhanced by reinforcing connections within modules or clusters. In contrast, when networks contain significant core-periphery structure, we find that learnability is best optimized by reinforcing peripheral edges between low-degree nodes. Overall, our findings suggest that the accuracy of human network learning can be systematically enhanced by targeted emphasis and de-emphasis of prescribed sectors of information.

Introduction

Whether we are learning a new language, navigating a social circle, or studying a textbook, our brains represent this information as a network of associations. However, human information processing is limited. We tend to "compress" these networks, leading to mental maps that are often blurry or inaccurate. This study asks a fundamental pedagogical question: If you want someone to learn a specific network, should you show it to them exactly as it is? Or, knowing how the human brain tends to fail, should you present a modified version that emphasizes the parts humans usually miss? This research moves beyond describing learning to optimizing how information is taught.

Methods

The researchers utilized a computational framework based on a maximum entropy model of human learning. This model posits that humans face a trade-off between accuracy and complexity: we try to learn the transitions in a network while minimizing the mental resources required.

The core approach involves:

  • The Learning Function ($f$): An analytic model that predicts the "learned" version of an input network based on a learner's accuracy parameter ($\beta$).
  • Optimization via KL Divergence: The Kullback-Leibler (KL) divergence was used to measure the "distance" between the predicted mental map and the true target network.
  • Dual Annealing: This global optimization algorithm was used to find the optimal input network ($A^*$) by minimizing the KL divergence.

The study evaluated this strategy across synthetic networks (modular, lattice, hierarchical) and real-world semantic networks extracted from 10 college-level linear algebra textbooks, where nodes represent mathematical concepts and edges represent their co-occurrence.

Optimization of a modular network. (A) Target modular structure. (C) KL divergence landscape for beta=0.3. (E) Optimal weights vs. learner accuracy.

Key Results

The "Exaggeration" Strategy

Much like animals use exaggerated signals in noisy environments, the optimal way to teach a network is to exaggerate its most difficult-to-learn features. In modular networks, the best strategy is to reinforce connections within clusters and de-emphasize the "bridge" edges between them. In core-periphery networks, where central nodes naturally dominate attention, the optimal strategy is to de-emphasize the core and heavily reinforce "peripheral" edges between low-degree nodes.

Centrality and Optimal Weight

There is a clear inverse relationship between an edge's betweenness centrality and its optimal weight. The more central an edge is to the network's flow, the more it should be de-emphasized to prevent it from washing out the rest of the structure.

Impact of Learner Accuracy

While specific optimal weights change depending on the learner's accuracy ($\beta$), the general strategy of emphasizing clusters and periphery remains effective. For highly accurate learners, the optimal input network eventually converges to the true target network.

Optimization landscape for a high-accuracy learner (beta=5).

Discussion

Surprisingly, in the early stages of learning, the human learning model—with all its biases—actually outperformed a mathematically "perfect" Maximum Likelihood Estimate. Our inductive biases toward clustering allow us to grasp the "big picture" of a network faster than a purely unbiased observer.

Future work should explore allowing the addition of entirely new edges to the input and testing these optimized distortions on actual students. A truly optimized educational tool would likely need to estimate a student's $\beta$ in real-time to tailor the material's emphasis, potentially improving long-term retention and conceptual understanding.

Summary updated

Abstract

Precisely how humans process relational patterns of information in knowledge, language, music, and society is not well understood. Prior work in the field of statistical learning has demonstrated that humans process such information by building internal models of the underlying network structure. However, these mental maps are often inaccurate due to limitations in human information processing. The existence of such limitations raises clear questions: Given a target network that one wishes for a human to learn, what network should one present to the human? Should one simply present the target network as-is, or should one emphasize certain parts of the network to proactively mitigate expected errors in learning? To investigate these questions, we study the optimization of network learnability in a computational model of human learning. Evaluating an array of synthetic and real-world networks, we find that learnability is enhanced by reinforcing connections within modules or clusters. In contrast, when networks contain significant core-periphery structure, we find that learnability is best optimized by reinforcing peripheral edges between low-degree nodes. Overall, our findings suggest that the accuracy of human network learning can be systematically enhanced by targeted emphasis and de-emphasis of prescribed sectors of information.
BibTeX
@inproceedings{qian2022optimizing,
  author = {William Qian and Christopher W. Lynn and Andrei A. Klishin and Jennifer Stiso and Nicolas Christianson and Dani S. Bassett},
  title = {{Optimizing the human learnability of abstract network representations}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2022},
  doi = {10.1073/pnas.2121338119},
}

Loading related content...

Figure