fly.ai / flybrain

The full 166,700-neuron MaleCNS graph runs as leaky integrate-and-fire units with a stated update rule (v <- exp(-dt/tau) v + gain * W @ spikes + tonic + noise, spike and reset at 1) and stated, explicitly hand-calibrated parameters, so membrane voltages change under a documented neuron model. Not L2, and the authors rule it out themselves: the Flybook market flies have a variable literally called dopamine with a learning rate, but the docstring states the learning is 'all at the interface between the market and its brain; the connectome itself is not rewired, because in this model the mushroom body, where flies really learn with dopamine, fires at its ceiling and can't learn yet.' The dopamine term writes only per-sense input gains and per-action output biases outside the graph - stimulation and a bolt-on policy, not synaptic plasticity. The connectome matrix is built once in the constructor and never written to again, and the sshfighter application trains a linear readout on descending neurons, the classic frozen-graph-plus-trained-readout pattern. The README is unusually honest about its own negative results (vision dies at the first relay; the 'real wiring matters' test fails at 2 ms).