flyverse

A carefully specified Shiu et al. 2024 LIF over the MaleCNS connectome - v_rest/v_reset -52 mV, v_th -45 mV, tau_m 20 ms, tau_syn 5 ms, 2.2 ms refractory, 1.8 ms conduction delay - with a graded rate model for the optic lobe. Squarely L1. Not L2: the README says outright that the connectome is never trained or edited, and the code backs it up. The nearest misses are both intrinsic or off: spike-frequency adaptation and a Tsodyks-Markram short-term depression term, and the latter ships disabled (std_u = 0.0) and in any case is a per-presynaptic-neuron resource that recovers in ~300 ms, not a stored change to a synapse. The only trained parameters in the repo live in a bolt-on 4-16-2 MLP motor decoder fitted with Adam, whose own docstring is 'Train a named motor boundary without changing connectome weights' - the classic trained-readout pattern, which is L1. No validation mark, but the near-miss deserves recording: the repository ships a 142-row expectation ledger scoring measured model responses against published fly electrophysiology and imaging — Maisak 2013 T4/T5 direction selectivity, Schnell 2010 HS optomotor, von Reyn 2014 giant fibre, Hallem & Carlson 2006 ORN rates, Seelig & Jayaraman 2015 EPG bump — one citation and criterion per row, scored pass or fail, published with its failures including an LC11 population showing the opposite size selectivity from the animal, and stated never to be a fitting target. The measurements it cites are peer-reviewed; the comparison is the project’s own, and the mark is reserved for comparisons published through peer review.