fly-brain
FlyWire LIF simulation with a natural-language interface and project-defined dopamine-learning experiments.
Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.
Dispute this rating →The evidence
No evidence, no level.The neuron model is leaky integrate-and-fire with documented parameters: -52 mV rest, -45 mV threshold, 20 ms membrane time constant, 5 ms synaptic time constant, 2.2 ms refractory period.
https://github.com/lixiang1076/fly-brain/blob/f6d525ab10c7a3ed4853d3d8b4b5e73f24cfe52b/code/paper-phil-drosophila/model.py#L24-L53Learned multipliers are written directly into the weights of the Brian2 Synapses object built from the FlyWire connectome, not into a separate readout layer.
https://github.com/lixiang1076/fly-brain/blob/f6d525ab10c7a3ed4853d3d8b4b5e73f24cfe52b/dopamine_learning.py#L199-L226Reward and punishment scale KC->MBON weights only for the Kenyon cells that actually spiked in the preceding simulation, with bounds of 0.01x to 10x baseline.
https://github.com/lixiang1076/fly-brain/blob/f6d525ab10c7a3ed4853d3d8b4b5e73f24cfe52b/dopamine_learning.py#L127-L146The dopamine teaching signal is a chat keyword applying a fixed strength, not spiking activity of the simulated PPL1 or PAM cells.
https://github.com/lixiang1076/fly-brain/blob/f6d525ab10c7a3ed4853d3d8b4b5e73f24cfe52b/fly_chat.py#L481-L501
Second reading, downgraded from a proposed L2. Real KC→MBON weights in the Brian2 object built from FlyWire do change, and the direction is canonical — punishment depresses, reward potentiates, on exactly the Kenyon cells that spiked. What is missing is the gate. The dopamine signal is not simulated at all: PPL1/PAM "activation" is a printed message, and the teaching signal is a keyword the user types applying a fixed multiplier. L2 asks for a rule gated by a modelled neuromodulator, and a constant behind a print statement is naming, not mechanism. The all-KC-to-all-MBON form also has no compartment structure, which is the part of the biology that makes reward and punishment target different MBONs. L1 is solid: a whole-brain LIF simulation with parameters from Shiu et al. and Kakaria & de Bivort.
Record
Nearby
All projects →Frozen language model coupled to the retained MaleCNS graph through a trained readout adapter. Language ability comes from the pretrained language model.
A MaleCNS simulation receives frames from insect videos on a virtual phone. Local Python/C++ simulation, browser observation chamber, and neural telemetry.
Ratings on thelearningfly.com are proposals, not verdicts. Every one of them is arguable in public.