fly-flappy-bird
MaleCNS wired to Flappy Bird with engineered adapters and recorded gameplay.
Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.
Dispute this rating →The evidence
No evidence, no level.A dopamine-gated three-factor rule writes into the connectome weight vector during the episode.
https://github.com/arjunkshah12345-hash/fly-flappy-bird/blob/f1a3ce657eae9f706e3e309cf64ee31f43bb5429/fly_brain.py#L213-L231The plastic edge mask is exactly the real KC->MBON edges of the released MaleCNS subgraph.
https://github.com/arjunkshah12345-hash/fly-flappy-bird/blob/f1a3ce657eae9f706e3e309cf64ee31f43bb5429/expand_real_brain.py#L143-L146Reward is fed to the plasticity rule every step inside the episode loop, so the update is online rather than an offline training pass.
https://github.com/arjunkshah12345-hash/fly-flappy-bird/blob/f1a3ce657eae9f706e3e309cf64ee31f43bb5429/train.py#L18-L31The recorded demo runs TunedBrain, whose step() has no plasticity - the shipped result is a frozen graph plus an 8-parameter evolved readout.
https://github.com/arjunkshah12345-hash/fly-flappy-bird/blob/f1a3ce657eae9f706e3e309cf64ee31f43bb5429/train_evo.py#L19-L48The authors report their own negative result: online dopamine-style plasticity on KC->MBON edges made play worse.
https://github.com/arjunkshah12345-hash/fly-flappy-bird/blob/f1a3ce657eae9f706e3e309cf64ee31f43bb5429/README.md#L7
A real MaleCNS v1.0 subgraph (3,082 cells, 120,830 released edges) driving leaky integrate-and-fire units, which earns L1. Not L2, for two reasons that compound. The default run has no plasticity: the README Quickstart runs record_play.py, which uses TunedBrain - a subclass whose step() drops the plasticity block - and then loads data/trained_weights.npz. That shipped weight array is bit-identical to the base connectome weights across all 120,830 edges, so the headline 50-pipe run changes zero synapses; what was trained is an 8-parameter evolved readout. The README says as much in its own first line: a motor readout tuned 'on top of frozen biological synapses'. And where plasticity does run, in FlyBrain.step via train.py, the gate is not a modelled neuromodulator: self.dopamine is the game reward accumulated and decayed (reinforce() adds reward * 0.8 and clips), while the 60 DAN cells that are present in the graph are never read. The authors are admirably direct about the limit - BLOG.md notes that 'neuromodulation is not a scalar dopamine knob', and the README reports the negative result that online dopamine-style plasticity on Kenyon-to-MBON edges made play worse. The plastic edge mask is genuine (pre in KC, post in MBON, computed from the released subgraph rather than synthesized), which is the strongest thing here and why this is a near-miss rather than a distant one. Not validated: no published comparison against real fly measurements.
Record
Nearby
All projects →MaleCNS simulation connected to ViZDoom through modeled visual inputs and a fixed button readout. Includes plasticity experiments and negative validation results; learned survival has not been demonstrated.
Research framework comparing a MaleCNS-constrained sparse recurrent controller with rewired graphs, conventional neural networks, and other controls on simple VizDoom tasks.
Ratings on thelearningfly.com are proposals, not verdicts. Every one of them is arguable in public.