Fruit fly connectome
Behavior pathways picked from the connectome of a fruit fly's whole central nervous system, distilled into a few dozen NAND gates. Not yet checked by independent experiments, and not on chain.
Status: two pathways are built as gate circuits, each bit for bit identical to its model. None of them has passed an independent biological experiment yet. Nothing is written on chain.
Data
We use the public connectome of the whole central nervous system of a male fruit fly (MaleCNS). All of it is downloaded:
| Item | Count |
|---|---|
| Neurons | 176,422 |
| Cell type groups | 11,752 |
| Connections between types | 3,853,300 |
| Synapses | 125,024,863 |
How it works
- Pick a behavior pathway with experimental support, such as "take off and escape when something looms".
- Use the connectome weights to build a simple firing-rate model as the teacher. The model's dynamics and signs are assumptions, not measured biological parameters.
- Calibrate the gains with looming stimuli and control stimuli: the target neuron must fire for looming and stay quiet for controls.
- Distill it into NAND gates on a finite table of 256 inputs, then prove the netlist matches the teacher bit for bit.
The two pathways built so far
| Pathway | NAND gates | Depth | Latches | Against the teacher | Calibration result |
|---|---|---|---|---|---|
| Giant fiber escape (GF) | 36 | 10 | 0 | 256 / 256 bits match | 102 positives; all 128 controls quiet |
| Visual descending (DN) | 16 | 6 | 0 | 512 / 512 bits match | 84 and 85 positives for the two outputs; all 128 controls quiet |
Joined into one three-part graph (plus a 6-gate collector), they come to 58 NAND gates and 406 bytes, with 768 / 768 bits matching. This is only a plan. It is not written on chain.
Not done yet
- Independent experimental checks: 0. So far there is only qualitative calibration and a logic equivalence check. The responses have not been reproduced on held-out biological experiment data.
- The input stimulus curves are assumed. They are not measured neural activity, and not rendered images.
- The encoder from raw senses to these inputs is not built yet.
- Pathways for courtship, odor approach and avoidance, navigation turning, and walking rhythm are still being modeled or are candidates.
- The logic equivalence proves "netlist = model", not "model = fruit fly".