flyboard

Every traced neuron of MaleCNS v1.0 (165,122 neurons, connections with 5+ synapses, signed by predicted neurotransmitter) is simulated as a leaky integrate-and-fire unit on GPU with parameters following Shiu et al. 2024 and one deliberate documented deviation (w_syn rescaled from 0.275 to 0.125 mV because MaleCNS neurons carry ~1.4x more synapses), and both the song scores and the command-mode body movements are read out of those simulated spikes. Not L2: the sparse weight matrix is built once in the Brain constructor and only ever appears inside torch.sparse.mm in the step loop - nothing is trained anywhere in the repo, not even a readout; the score is a cosine similarity against the response to a reference courtship song, which involves no fitting at all. No validation mark either, despite the unusually strong self-criticism: the ear-only ablation, the shuffled-session robustness checks and the command-neuron-to-motor-output table are internal consistency checks, not a published comparison against real fly measurements. The project's own honest reading - that this fly mostly hears the balance between its two ear channels, barely distinguishes a lover from a wasp, and that they switched from mean to median after seeing a motor circuit latch on - is the most useful thing on the page and should carry over into any summary.