Zebra Fish

A larval zebrafish with 71,721 recorded neurons

0x94b5d6857c39f8a5c77c6062bcce79d628f7a124

[Zebra Fish @ZebraFishlink]

A Zebrafish's vertebrate brain, wired, recorded, and put to work.

In 2024 a larval zebrafish sat in a virtual environment while a light-sheet microscope recorded its whole brain: 71,721 neurons, 7,879 frames, roughly one volume per second, through nine stimulus conditions. Google Research and Janelia released that recording as ZAPBench, a benchmark whose question is simple to state: given the recent past of every neuron, can a model forecast the next 32 frames?

ZEBRALINK is what happens when you take that recording seriously as a thing you can build on.

What we have. The public ZAPBench release: the traces, the segmentation, the condition boundaries. We download a checksum-locked subset, 8,192 cells across sixteen regions of the brain, plus the measured positions of all 71,721. Every number on our pages can be traced back to a row in those files.

What we built. A linear model of 2,048 recorded cells: three frames of history, ridge regression, nothing exotic. Its one-step fit on the training conditions is R² 0.92. Rolled forward on the condition it never saw, it drifts within about ten frames. That is not a criticism. That is what a linear model of a brain looks like, and it is exactly the kind of baseline ZAPBench exists to beat.

We then chose cells from the recording itself, before any task. 256 cells that respond to full-field flashes: the eyes, as far as this model has eyes. 128 cells on the left of the hindbrain and 128 on the right that fire during turning: the hands. 16 gate cells. 15 and 2 cells to receive reward and its opposite, mirroring in count only the dopamine cells the fruit-fly project used. No atlas. No retinotopy. The positions are ranks and fitted weights, and we say so on the page.

What it does. A 320×180 view of the world is sampled at the light cells. Right minus left hindbrain becomes a control signal. The gate cells decide when to act. With those three rules the same brain holds a table-tennis paddle, a set of handlebars, and a leash, live in your browser at 16 Hz, with the receptors that fire drawn on screen as they fire.

The results are what they are.

Table tennis: about one return in three over a minute.

Bike: around a hundred metres per half-minute with training wheels on, and a fall without them.

Dog: on the path about half the time; squirrels are the failure mode. We ran the learning rule, an anti-Hebbian eligibility model borrowed from STONKFLY and Huang, Luo et al. 2024, on 64,748 fitted edges, then froze it and ran again.

Opensource : https://t.co/poRrh099aM

Web simulations : https://t.co/kmUyBVpPkg

[Zebra Fish @ZebraFishlink]

Nothing much, just a zebrafish launching a coin on @ponsdotfamily

[Zebra Fish @ZebraFishlink]

The fish launched Zebra Fish.

Ticker $ZEBRA · Pons · paired with GOOGL

CA: 0x94b5d6857c39f8a5c77c6062bcce79d628f7a124

Code and docs: https://t.co/HxPSuftQIt

Data: ZAPBench, Google Research + Janelia