Fly / Wirehead
A MaleCNS simulation receives frames from insect videos on a virtual phone. Local Python/C++ simulation, browser observation chamber, and neural telemetry.
Synapses change under a modeled rule grounded in fly biology — for example dopamine-gated change on Kenyon-cell → MBON synapses in the mushroom body. Learning begins here.
Dispute this rating →The evidence
No evidence, no level.The plasticity rule is a baseline-centered anti-Hebbian eligibility model adapted from Huang, Luo et al. 2024, with declared trace constants and efficacy bounds, taking KC and DAN firing rates as its only drive.
https://github.com/mattyhempstead/fly-wirehead/blob/fcefe9441f80e25aab713411ebced53f5e5ea172/flywirehead/neural/rule.py#L1-L65Each simulation bin writes the updated efficacies back into the connectome's own weight array, and this runs whenever weights are not explicitly frozen.
https://github.com/mattyhempstead/fly-wirehead/blob/fcefe9441f80e25aab713411ebced53f5e5ea172/flywirehead/neural/brain.py#L331-L352The plastic edges are Kenyon-cell inputs onto MBON07 and MBON11, in two compartments driven by the 15 PAM11 and 2 PPL101 cells, with modulation gains derived from anatomical DAN-to-MBON contact counts.
https://github.com/mattyhempstead/fly-wirehead/blob/fcefe9441f80e25aab713411ebced53f5e5ea172/flywirehead/neural/circuit.py#L1-L45Plasticity is enabled by default: the engine passes learning=not weights_frozen, and weights_frozen initializes to False.
https://github.com/mattyhempstead/fly-wirehead/blob/fcefe9441f80e25aab713411ebced53f5e5ea172/flywirehead/engine.py#L60-L70The authors describe the observed weight change as a mechanism check and explicitly disclaim any claim of learned preference or biological validity.
https://github.com/mattyhempstead/fly-wirehead/blob/fcefe9441f80e25aab713411ebced53f5e5ea172/docs/validation.md#L1-L12
A spiking LIF simulation of the full retained MaleCNS v1.0 graph (166,700 neurons, 25.6M edges) whose plastic edges are the KC inputs onto MBON07 and MBON11, with PAM11 and PPL101 as the two dopamine compartments and modulation distributed by the anatomical DAN-to-MBON contact fractions. During each observation the rule writes new efficacies into the live connectome weight array, and plasticity is on by default (weights_frozen defaults to False, and the engine passes learning=not weights_frozen). The rule is an adaptation of the published Huang/Luo 2024 baseline-centered anti-Hebbian rate rule, driven by actual full-network spike counts rather than by an optimizer gradient - there is no torch optimizer or trained readout anywhere in the project. No validation mark: the project's own validation log describes its plasticity result as a mechanism check and states explicitly that it is not a claim of learned preference or biological validity, and no comparison against real fly measurements is published.
Record
Nearby
All projects →Ratings on thelearningfly.com are proposals, not verdicts. Every one of them is arguable in public.