fly-brain-bench

Leaky integrate-and-fire over real FlyWire data shipped with the page, using the Shiu et al. 2024 constants and a closed-form two-variable integration, with an event-driven active set - a documented neuron model, so L1. Not L2: the simulation never writes a weight. The in-browser 'teaching' feature fits a multinomial logistic regression over 58 descending/motor cell-type rates by gradient descent, and the README states in its own words that this trains the readout, not the brain, and that no synapse changes. The repo does contain a mushroom-body plasticity experiment (tools/19_isolated_mb.py) that depresses real KC->MBON edges to 25% for the Kenyon cells active during the conditioned odour and re-probes; I read it as an offline one-shot perturbation inside an analysis script - no dopamine signal, no update during a run, no effect on the shipped simulation - rather than a plasticity rule the model runs, which is why it does not lift this to L2. The author's negative result from that script is the most valuable thing on the page: the model's Kenyon-cell code is 60.7% active with 99.3% overlap across different odours, so dopaminergic depression cannot produce odour-specific learning here.