Work · Experimental substrate
The fly connectome as a testbed, not a proof
I test my hypotheses on networks modelled from a wiring topology derived from the fruit fly connectome (FlyWire). It is a source of structure and a falsification bench. It is neither the aim of the project nor automatic evidence of biological fidelity — and this page explains why the distinction matters.
A testbed, not a definition
The fruit fly connectome — the map of its neurons and their connections, 139,255 neurons in version 783 published by FlyWire — is one of the few complete wiring maps of a real organism. I use it as a starting structure: a topology on which to place model neurons, rather than an invented graph. It is at this complete scale that I checked the replay of the numerical engine (see below); the learning prototype itself runs on a much smaller circuit, modelled from part of this connectome.
But my work is not defined by the fly. The architecture must be able to move to other connectomes, other sensors, other morphologies without losing its learning mechanisms. The connectome is a foothold and a falsification bench — ground on which hypotheses can be proved wrong — not the purpose. The embedded AI I am aiming for will reuse mechanisms understood on this testbed, not these data.
Determinism holds at full scale
A separate question: does the computational substrate itself — the ArithSpec/Q16.16 integer arithmetic on which my models run — remain bit-exact as the graph grows to its real size? This is a numerical engine test, not an execution of the learning prototype; the distinction matters, see below.
Demonstrated
What this test establishes — and what it does not
- It establishes that the arithmetic engine remains reproducible bit for bit as graph size increases to the real scale of the connectome — a result about the substrate, independent of any task.
- It does not establish that my models run the complete connectome. The learning prototype (see the roadmap) runs on a much smaller circuit, a typed subgraph modelled from part of this connectome. The simulated quadruped organism is a separate system: during its four-hour run, its controller executed a 50,000-neuron subgraph drawn from this connectome (see the page of the simulated organism).
- It establishes no peak-memory figure for a full-scale execution — that figure has not yet been measured reliably and is therefore not advanced here. The 144 MB shown on the organism page measures the four-hour run of the simulated organism (50,000-neuron subgraph) — not a full-scale execution.
Topology source: FlyWire Whole-brain Connectome Connectivity Data, version 783.0 (FlyWire Consortium / Dorkenwald et al., 2024, “Neuronal wiring diagram of an adult brain”, Nature 634), DOI 10.5281/zenodo.10676866, CC BY 4.0 licence, open access — with the same reservations as everywhere else on this page: real structure, no biological function claimed.
The vocabulary I rule out
I apply a discipline of language throughout this site:
| Never | Instead |
|---|---|
| “biological connectome” | “candidate topology derived from FlyWire” |
| “biological synaptic weights” | “edge counts derived from reconstruction” |
| “the real mushroom body” | “structural subgraph, typing to be revalidated” |
| attributing the function to the connectome | — not established; only after beating a rewiring control |
| cognition, evolutionary conservation, fidelity to living systems | never, under RG0 scope |
Scope of the result
- The topology is derived from FlyWire, not a biological connectome. The weights are edge counts from reconstruction, not measured conductances.
- No biological function, no cognition, no evolutionary conservation and no generalisation to living systems is claimed.
- Attribution of substructures requires rigorous typing before any biological statement.
- The neural dynamics, the weights and the learning mechanisms belong to the project — the connectome supplies only the wiring structure.
- See also the quantitative results on this topology in Paper E, and its use in the simulated quadruped organism.
The connectome supplies the structure; the dynamics, the weights and the learning are those of the project.