Paper Overview
Field: Machine Learning Author: Yeyuan Liu Date: 2026-11-10 arXiv: 2611.14508
Summary
Connectome-constrained models are often used to argue that a measured wiring diagram outperforms shuffled or rewired wiring, and hence that structure shapes function. Such comparisons can confound wiring with how the models are tuned: a connectome fixes who connects to whom, not the excitability, gains or plasticity rates the model also needs, and modellers often tune these until the real-wiring model works.
The paper separates three estimands:
- P1: physiology tuned on the real wiring and transferred to nulls
- P2: each wiring tuned on its own
- P3: P2 restricted to physiology consistent with independent data
- Robust conclusions (hold under all protocols): segregated reward/punishment dopamine helps learning; head-direction topology is necessary to maintain persistent activity wave peaks.
- Protocol-dependent conclusions: real Kenyon cell output connectivity in the mushroom body beats cross-subtype-identity shuffles under P1/P2 (margin 0.09), but loses by 0.09 under P3 (95% CI excludes zero; held-out runs confirm). The verdict is determined by the fraction of alpha'-beta' cells among active Kenyon cells—the fraction selected by P1 is too low. At the proportion implied by imaging data (~44%), the real-wiring model fails to learn, exposing model incompleteness rather than a defect in the fly connectome.
- In flyvis, P1 penalizes the mirrored connectome more than spatial shuffling (for both 50/50 member types); rankings persist after 10x retraining, and depend on checkpoint choice at 50x.
The biases involved are known in machine learning (under-tuned baselines, inherited weights, winner's curse). The author measures their effect on biological conclusions using synthetic ground truths, placebo null models, and three Drosophila loop models.
Findings
Conclusion
Credible structural claims require matched optimization, physiology constrained by independent data, and models that still work under that constraint.