FlyScale: geometric scaling of the adult Drosophila connectome

Can a biological connectome be enlarged? We take the complete adult Drosophila connectome (FlyWire FAFB v783; 139,255 neurons, 2,700,513 thresholded connections, 34,153,566 synapses) and ask whether a synthetic graph can be generated that preserves its structural statistics, its propagation dynamics, its behavioral capabilities and its energy budget as it grows.

The method is a geometric one. We first fit latent coordinates to the connectome and ask which space actually predicts connectivity: a 32-dimensional spectral embedding predicts held-out connections best (AUC 0.952), a 2-D hyperbolic embedding beats the anatomical 3-D soma coordinates (0.865 vs 0.851), and no embedding beats a degree-only baseline by a margin larger than its own run-to-run spread. Scaling then proceeds by node subdivision in that geometry: each neuron is replaced by c children placed near it, each inheriting the parent's partner set and synapse weights, with the concrete target chosen by the fitted connection law P(connect | distance) = 1/(1+e^((d-R)/T)).

The result is a ladder of graphs at 2×, 5× and 10× (up to 1,392,550 neurons and 27.8M connections) whose connection count scales as N^1.0175 (R^2=0.9995) — sparsity is preserved, mean degree and mean connection strength stay bounded — and which renormalizes back to the biological graph almost exactly. Capability grows on some axes (stimulus-discrimination information ∝ N^1.014, memory capacity ≥ 96 associations versus 8 at 1×) and is flat on others. Energy is measured on two tracks that are never summed: hardware (438/500/946/2022 GPU joules across the ladder, with cost per spike falling from 8.9 to 4.2 μJ) and biological-equivalent.

The work is deliberately dual-implementation: a Python reference oracle and a Vyb-native production path — loader, discrete-event runtime, GPU kernels — and the effort to build the second one surfaced twelve compiler defects, of which the blocking one, a kernel-mode store that wrote eight bytes into a four-byte slot, is documented, fixed upstream and verified here. We report what failed as carefully as what worked.