Kenyon · male CNS connectome · Janelia + Google Research

We taught a fruit fly to make beautiful websites

2,045 measured Kenyon cells, a measured feedback loop and 17,985 measured synapses — and the ones firing right now are turning inside its head. It was shown two versions of the same Ammoura storefront, identical words and opposite design, then rewarded on the one built from Ammoura's own theme tokens and punished on the one built the way generated design usually is. Everything on the screen below is what it picked.

—Kenyon cells

drag its head to turn the circuit · scroll to zoom

Kenyon cells —

avoid0.000approach
approach twin generation 00 / 15
gen 00

What the fly thinks of this page

+0.000approach − avoid

Its preference for the best candidate, generation by generation. A page it has no opinion about scores zero. For scale, an odour the fly was trained on directly reaches ±0.25.

The two brains, after the lesson

Synaptic efficacy of the approach twin
taught to like · — depressed
Synaptic efficacy of the avoid twin
taught to avoid · — depressed

2,045 × 49 synapses, blocked down the long side. Lit means the lesson weakened a synapse there; black means it never touched one. This is the only difference between the two flies.

What each fly ended up building — and what holds up

Six flies ran from the same starting pages with the same mutations: the two twins, a naive fly with no lesson, and three mis-aimed flies given the same reward with the dopamine delivered to the wrong compartments. KC overlap is how much of the Kenyon-cell code each final page shares with the page the fly was taught, and vs random places that against 30 pages nobody selected.

The write-up

What this actually taught us — including the parts that did not work, and an honest comparison against the ordinary ways of getting the same result. It is being written now. Leave an address and it arrives once, as a PDF, when it is finished. A draft is readable now.

One email, the report, nothing else. The address is stored and not shared.

What this is standing on

None of the wiring here was invented. It was measured by other people, over years, and published so that anyone could run it. These are the specific things this page is built out of.

The animal

The circuit

  • Kenyon, 1896 The brain of the bee. The cells this project is named after, drawn through a microscope 130 years ago. 10.1002/cne.910060302
  • Caron, Ruta, Abbott & Axel, 2013 Olfactory input to Kenyon cells converges at random — why the expansion layer can be treated as an arbitrary code. 10.1038/nature12063
  • Lin, Bygrave, de Calignon, Lee & Miesenböck, 2014 APL's feedback makes the code sparse. The 8% this model reproduces is their result. 10.1038/nn.3660
  • Aso et al., 2014 The compartment map: which dopamine neuron teaches which output neuron, and what each compartment means. 10.7554/eLife.04577
  • Hige, Aso, Modi, Rubin & Turner, 2015 The plasticity rule itself — dopamine depresses the KC→MBON synapses that were active. It is the only thing that changes here. 10.1016/j.neuron.2015.11.003
  • Dasgupta, Stevens & Navlakha, 2017 The same circuit read as an algorithm. The reason a page can be handed to it as if it were a smell. 10.1126/science.aam9868

The build

The rules it is judged by

  • Ammoura the storefront this is trained on. The tokens are read straight out of its src/app.css by ammoura.py, so the reference moves when the product does. The page it was rewarded on is built from them; the one it was punished on is not.
  • NebulaKit where the contrast gate came from — a design system that fails a build over it.
  • WCAG 2.1 contrast minimum 4.5:1 for normal text, 3:1 for large. Measured on every page here, not asserted.
  • *Space · Ammoura · Guides