Embodied fly-brain
Community research repository combining a FlyWire spiking model with NeuroMechFly/MuJoCo, sensory experiments, and an accompanying preprint.
Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.
Dispute this rating →The evidence
No evidence, no level.Plasticity is initialised over a mutable view into the connectome's sparse CSR weight values, with sign preserved and per-synapse clamp bounds.
https://github.com/erojasoficial-byte/fly-brain/blob/27cec28d5d202eb004683fb4c1a1033eec8deea0/brain_body_bridge.py#L337-L383The Hebbian update writes dW = eta * pre * post * sign - alpha * W into the connectome weights in place.
https://github.com/erojasoficial-byte/fly-brain/blob/27cec28d5d202eb004683fb4c1a1033eec8deea0/brain_body_bridge.py#L385-L399The plasticity update runs inside the per-timestep brain step, every 10 steps, under torch.no_grad - not in an offline training script.
https://github.com/erojasoficial-byte/fly-brain/blob/27cec28d5d202eb004683fb4c1a1033eec8deea0/brain_body_bridge.py#L446-L461The neuron model is AlphaLIF with documented synaptic and membrane time constants, threshold and refractory period.
https://github.com/erojasoficial-byte/fly-brain/blob/27cec28d5d202eb004683fb4c1a1033eec8deea0/code/run_pytorch.py#L35-L48The preprint's headline result compares two simulated flies to each other, not to real fly measurements.
https://github.com/erojasoficial-byte/fly-brain/blob/27cec28d5d202eb004683fb4c1a1033eec8deea0/README.md#L41
Second reading, downgraded from a proposed L2. The rule runs in-simulation and writes into a mutable view of the FlyWire CSR, with no optimizer in the loop — all true. But it is dW = eta * pre * post * sign − alpha * W applied across all ~15M synapses: textbook Hebbian learning with weight decay, not a fly mechanism. L2 asks for a rule grounded in fly biology — a modelled neuromodulator, a compartment, a timing window — and there is none here: no dopamine term, no KC→MBON, no cell type in the update at all. L1 is well earned on the AlphaLIF dynamics (5 ms synaptic and 20 ms membrane time constants, threshold, refractory period) over the FlyWire graph, driving a NeuroMechFly body. The accompanying preprint compares two simulated flies against each other rather than against real ones, so no validation mark either.
Record
Nearby
All projects →Research code for the Shiu et al. connectome-based leaky integrate-and-fire model, including activation/silencing experiments, notebooks, and FlyWire data configuration.
FlyWire whole-brain LIF implementation based on Shiu et al., with multiple simulation backends and benchmarking tools. This repository supplies the neural model; it is not a complete embodied demo package.
Ratings on thelearningfly.com are proposals, not verdicts. Every one of them is arguable in public.