Iterated learning: Intergenerational knowledge transmission reveals inductive biases

Michael L. Kalish, Thomas L. Griffiths, Stephan LewandowskyPsychonomic 2007 · DOI

Introduction

How does knowledge change as it passes from person to person over generations? This process, known as "iterated learning," is a fundamental mechanism of cultural transmission. Whether it is the evolution of language, the persistence of social norms, or the spread of religious beliefs, information is rarely learned from a perfect source. Instead, we learn from individuals who themselves learned from a previous generation.

A central question in cognitive science is what determines the ultimate form of this transmitted knowledge. Does it eventually degrade into random noise, or does it settle into a stable state? This research demonstrates that iterated learning acts as a powerful filter, stripping away specific data to reveal the underlying inductive biases—the default assumptions—of the human mind.

The Bayesian Framework of Iterated Learning

The authors formalize iterated learning using a Bayesian framework. In this model, each learner is viewed as a Bayesian agent who uses a prior distribution (representing their innate or internalized biases) and a likelihood (the data they observe) to form a hypothesis. This hypothesis then generates the data for the next learner in the chain.

Mathematically, this process can be described as a Markov chain. The researchers proved that a chain of such learners will eventually converge to a "stationary distribution" that is identical to the learners' prior. In other words, over many generations, the specific data provided to the first learner becomes irrelevant, and the information inevitably reflects the learners' inductive biases.

Iterated learning formalized as a Bayesian Markov chain, where data ($d$) informs hypotheses ($h$), which in turn generate new data.

Experimental Methodology: Function Learning

To test this theoretical model, the researchers conducted a behavioral experiment using a function learning task. They organized 288 undergraduate students into "families" or chains of nine generations.

The task involved learning a relationship between an input $x$ (the width of a bar) and an output $y$ (the height of a bar). The first participant in each chain was trained on one of four initial function types:

  1. Positive linear: $y = x$
  2. Negative linear: $y = -x$
  3. Nonlinear: A U-shaped quadratic function.
  4. Random: Completely uncorrelated $x,y$ pairings.

After the training phase, participants were tested on new $x$ values. Their predictions then served as the training data for the next person in the chain. This allowed the researchers to observe how the "knowledge" of the function evolved over nine generations.

Key Results: The Power of the Positive Linear Bias

The results provided striking evidence for the convergence hypothesis. Regardless of the starting condition, the vast majority of chains (28 out of 32) converged to a linear function with a positive slope.

Evolution of function learning across generations (A-E) and the corresponding change in correlation (F), showing convergence from various initial states to a positive linear relationship.

As shown in the figure above, the "positive linear" condition remained stable throughout the experiment. However, chains starting with negative linear, nonlinear, or random data rapidly shifted toward a positive linear relationship. This shift was often clearly visible within just three to five generations. Quantitatively, the correlation between the transmitted data and a positive linear model systematically increased over time for almost all conditions.

While the positive linear bias was dominant, there were a few exceptions. Three chains converged to a negative linear function, suggesting that while humans have a strong preference for positive correlations, our inductive biases are complex and can occasionally settle into other stable, though less common, states.

Insights and Implications

The study offers several profound insights into human cognition and cultural evolution:

  • Information Loss as a Feature: While we typically view the "Telephone" effect as a source of error, here it is revealed as a systematic process. Iteration "washes out" the noise of specific starting data, leaving behind the fundamental "shape" of the human mind.
  • Rapid Convergence: The speed at which these biases manifest suggests they are deeply rooted. You do not need hundreds of generations for cultural evolution to reveal underlying cognitive structures.
  • A Tool for Discovery: Iterated learning provides a "window" into hidden biases. It offers a method for researchers to uncover "default" assumptions in complex tasks where direct questioning of participants might be ineffective.

Conclusion and Future Directions

This research demonstrates that the structure of cultural information is not just a reflection of the world, but a reflection of the learners themselves. While this study focused on simple one-dimensional functions, the principles of iterated learning likely apply to far more complex domains, such as language structure or social hierarchies.

Future work remains to determine whether these inductive biases are innate (hard-wired) or the result of general-purpose learning from a world that frequently exhibits linear patterns. Regardless of their origin, these biases act as the ultimate destination for knowledge as it travels through the generations.

Summary updated

BibTeX
@inproceedings{kalish2007iterated,
  author = {Michael L. Kalish and Thomas L. Griffiths and Stephan Lewandowsky},
  title = {{Iterated learning: Intergenerational knowledge transmission reveals inductive biases}},
  booktitle = {Psychonomic},
  year = {2007},
  doi = {10.3758/bf03194066},
}
Figure