train-your-fly / connectome

The FlyWire v783 graph is used as fixed wiring for a PyTorch Geometric classifier, with an anatomically careful eye model in front and a Kenyon-cell linear readout behind. No neuron is simulated over time: `forward` runs exactly NUM_CONNECTOME_PASSES (3) message-passing steps, each computed purely from incoming input, and the repository contains no membrane variable, no time constant, no leak, no spike threshold and no reset - a grep for membrane/voltage/tau/refractory/spike across `src/` returns only the string 'leaky_relu'. The authors say this themselves: 'Neurons keep no state between steps: each step is computed from incoming input alone.' That absence of any state variable is what rules out L1; the per-neuron xi is a static learned bias inside the nonlinearity, not a spiking threshold. Not L2 either, and not even close: the connectome edge weights are a non-trainable `register_buffer`, what is learned is a separate `edge_weight_multiplier` gain per synapse plus an optional per-neuron threshold plus the final linear classifier, all optimised offline by AdamW on a cross-entropy loss. This is a careful and honest project - the README's 'What gets trained' table states the regime precisely - it just sits at the fixed-wiring end of the scale.