Two runners on the same track. One starts halfway down the course, the other from the starting line. Who's faster? The answer is obvious, but in brain science, such a trick went unnoticed for a long time. Neural networks that replicate the connections of a real fruit fly brain were considered better learners than random networks. They were praised for their quick start and time savings.
But now the runners were placed on the same line. Researchers took the connectome of a fruit fly and compared it with two types of random networks: a simple random one and a more sophisticated one where each "neuron" has the same number of connections as in the real brain. Under weak control (when both networks started from an already trained state, and the random network only roughly matched the brain in overall numbers) the biological wiring won. It reduced error faster, showed higher activity, and ran quicker.
When both networks started training from scratch, the speed advantage vanished. And when the random network was replaced with one that had the same distribution of connections, the apparent activity advantage also disappeared. Five different random networks with preserved connection counts confirmed: there is no difference. Earlier claims of biological topology superiority now appear to have been due to experimental conditions, not the wiring itself.
So the runner with a false handicap turned out to be just an ordinary runner. The fly brain's wiring may be good for a fly, but its advantage over random networks in learning tasks remains unproven. A fair start put everything in its place.
as written by the authors arXiv:2604.04033
Connectome-constrained neural networks are often evaluated against sparse random controls and then interpreted as evidence that biological graph topology improves learning efficiency. We revisit that claim in a controlled flyvis-based study using a Drosophila connectome, a naive self-loop-matched random graph, and a degree-preserving rewired null. Under weak controls, in which both models were recovered from a connectome-trained checkpoint and the null matched only global graph counts, the connectome appeared substantially better in early loss, mean activity, and runtime. That picture changed under stricter controls. Training both graphs from a shared random initialization removed the early loss advantage, and replacing the naive null by a degree-preserving null removed the apparent activity advantage. A five-sample degree-preserving ensemble and a pre-training activity-scale diagnostic further strengthened this revised interpretation. We also report a descriptive mechanism analysis of the earlier weak-control comparison, but we treat it as behavioral characterization rather than proof of causal superiority. We show that previously reported topology advantages in connectome-constrained neural networks can arise from initialization and null-model confounds, and largely disappear under fair from-scratch initialization and degree-preserving controls.
A fruit fly has about 130,000 neurons, and the map of their connections, the connectome, takes up gigabytes of data.