Gibbs sampling with people

Using collective human behavior to implement probabilistic inference and generative modeling.

This project investigates how collective human behavior can implement powerful computational algorithms for probabilistic inference. Inspired by Gibbs sampling—a core method in Bayesian statistics and machine learning—we design experiments in which individuals iteratively modify shared cultural artifacts, allowing group-level dynamics to approximate sampling from complex probability distributions.

By structuring social interaction as a computational process, we show that populations of human participants can collectively perform inference in high-dimensional generative spaces, revealing deep connections between cognition, social dynamics, and probabilistic computation.

Human Gibbs sampling paradigm

Our experiments transform creative human tasks—such as iteratively modifying images, melodies, or visual patterns—into distributed sampling procedures. This allows us to probe how mental representations, inductive biases, and cultural transmission shape the structure of learned distributions, bridging perception, learning, and collective intelligence.

What we do

  • Design large-scale iterated learning experiments that implement probabilistic sampling through human interaction.
  • Use generative modeling to reconstruct latent cognitive representations from behavioral data.
  • Study how individual inductive biases accumulate into collective representations.
  • Develop computational frameworks linking Bayesian inference, cultural evolution, and collective intelligence.
  • Explore applications to creativity, perception, and human–AI collaboration.
Iterated sampling and representation learning

Why it matters

Many core cognitive and cultural phenomena—from perception to creativity—can be understood as forms of probabilistic inference. This project demonstrates that groups of humans can collectively implement sophisticated computational algorithms, providing a new framework for studying cognition as a distributed process. Beyond theoretical insight, this work opens new pathways for designing hybrid human–AI systems, collective creative platforms, and large-scale cognitive experiments that harness social interaction as a computational resource.

Collective inference and emergent structure
  • placeholder: Gibbs sampling with people paper
  • placeholder: generative modeling methods
  • placeholder: experiment code and datasets