Does the perturbation budget transfer? Silencing interactions in a whole-brain spiking model of the fly are confined to a few strong cell types
Year: 2026
Type: preprint
Abstract: Paper 13 measured, in 50 task-trained models of the fly visual system, how silencing interactions scale with the number of cell types silenced, and turned the result into a budget: the order of battery that certifies untested combinations to a stated tolerance. Whether that shape belongs to connectome-constrained models in general or to one ensemble was left open. Here the same nested battery (orders 1-5, exact Möbius inversion, 1,860 sets, 74,440 trials) was run on a different model class — the whole-brain leaky integrate-and-fire model of Shiu et al. (2024), one neuron per Flywire neuron, one free weight, no training — on its sugar-sensing to proboscis-motor circuit, under a protocol frozen with Paper 13's published constants as cross-class predictions and a pilot disclosed as exposure. Spiking noise gave every interaction a standard error, and every energy was corrected for its noise floor. None of the four predictions held. For random combinations of the 131 responsive cell types, interaction energy at orders 2-4 was indistinguishable from zero (0.00002, −0.00002, 0.00002 against single effects of 0.0107) and at least five to thirty times below flyvis at orders 2-3; shares of combinations exceeding a 20% tolerance were zero to within noise; synapse-count anatomy ranked triple interactions weakly (Spearman 0.25) and quadruples not at all. Interactions were confined to a few strong types, where they were large (the two strongest nearly cancel each other, −0.56), decayed twelvefold from singles to pairs and threefold to triples, and then fell below the floor. The budget arithmetic therefore says single-type silencings certify combinations up to five types for this circuit, where flyvis needed triples. The perturbation budget is a measurement with a transferable method, not a law with transferable constants; the order a validation battery must close depends on the model and the circuit, and can be measured in advance. A bias in the preregistered bootstrap, found after the analysis, is reported and corrected.
Author: Ryan, Thomas
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225983392
