← All projects Instrumentation · Collective behaviour

Closed-loop projection for collective perception

Can a group detect a signal that its individual members can’t? I built a closed-loop projector system that tracks a whole school of fish and shows it moving visual stimuli in real time, to study how groups perceive and decide.

Where
Max Planck Institute of Animal Behavior · Konstanz, Germany
When
2016 – 2020
Role
Postdoctoral Researcher (with Iain Couzin)
Tools
Real-time tracking, projection, Python, Bayesian inference

The question

Animals in groups interact constantly, and these interactions give rise to collective behaviour that responds to the environment. How much information processing does a group actually do, and how does that depend on the size of the group? Answering this needs precise control over what every animal sees.

The build

  • Designed and built a large experimental facility that projects moving patterns onto the floor of a shallow aquarium
  • Tracked the fish in real time and closed the loop, so the stimulus could depend on where the animals were
  • Adapted a classic neuroscience perception task, the random-dot kinetogram, so whole schools could perform it. The school has to detect coherent motion hidden among random motion
The experimental facility at the Max Planck Institute of Animal Behavior.
The experimental facility at the Max Planck Institute of Animal Behavior.
Inside the arena: a school of fish swimming over moving dot patterns projected onto the floor of the tank.
Inside the arena: a school of fish swimming over moving dot patterns projected onto the floor of the tank.

The science

I turned social interactions into causal graphs of collective decision-making. With Bayesian modelling experts, I inferred how group size affects the group’s ability to process information, and designed perturbation experiments with noisy signals to test the models against real data.

Steering the school: dynamic pattern projection manipulates the direction in which the school mills.
Steering the school: dynamic pattern projection manipulates the direction in which the school mills.
Schools respond faster and more strongly as the stimulus becomes more coherent.
Schools respond faster and more strongly as the stimulus becomes more coherent.

Outcome

Manuscript in preparation: Bath, Graving, Walter & Couzin, “Causal role of group size in collective signal detection.” Invited talks at the International Centre for Theoretical Physics (Trieste, 2018), HHMI Janelia (2018) and the ASAB Summer Conference (Konstanz, 2019).

My contribution

I conceived the experiments, designed and built the facility, and ran the study. I also supervised a Master’s student who developed machine-learning algorithms to control fish schools in the arena.

Videos

Methods: how the experiment works.
Results: what the schools did.
The stimulus at 100% coherence: easy.
The stimulus at 50% coherence: harder.
Tracking every fish in the school (tracking by Tristan Walter).
A school reversing direction, with the distribution of headings below.
A school of 128 fish across trials of varying coherence.
Reversals across group sizes.