fly-hero
Frozen fly-connectome reservoir that plays Clone Hero by seeing the highway.
Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.
Dispute this rating →The evidence
No evidence, no level.The shipped graph is the larval Drosophila connectome from Winding et al. 2023, held frozen as the reservoir W.
https://github.com/actuallyrizzn/fly-hero/blob/0a058bb8ad1e6a8e84202f29cc5a96bf960d2ef1/src/flyhero/data/README.md#L1-L18Dynamics are a leaky tanh echo state over the connectome matrix, with stated leak and spectral radius; nothing here has a gradient.
https://github.com/actuallyrizzn/fly-hero/blob/0a058bb8ad1e6a8e84202f29cc5a96bf960d2ef1/src/flyhero/reservoir.py#L52-L59The state update itself: leaky tanh over W_in u + W x.
https://github.com/actuallyrizzn/fly-hero/blob/0a058bb8ad1e6a8e84202f29cc5a96bf960d2ef1/src/flyhero/reservoir.py#L116-L124In the live prosthesis loop the action is decided before the reservoir steps; the reservoir runs only for the visualiser.
https://github.com/actuallyrizzn/fly-hero/blob/0a058bb8ad1e6a8e84202f29cc5a96bf960d2ef1/tools/play_prosthesis.py#L853-L861The default play path substitutes NullReservoir, a pass-through test double.
https://github.com/actuallyrizzn/fly-hero/blob/0a058bb8ad1e6a8e84202f29cc5a96bf960d2ef1/src/flyhero/play.py#L24-L37
A real connectome is used - the complete larval Drosophila brain from Winding et al. 2023 (2,956 neurons, 116,922 edges) ships in the repo - and it is run as a leaky tanh echo state, x <- (1-a)x + a*tanh(W_in u + W x), with stated leak and a spectral radius rescale. Documented simplified dynamics, so L1. Not L2: the module is called 'Frozen recurrent graph, W does not train'; only a linear readout is fitted, by least squares, offline. A caveat a reader should see before treating this as a fly playing Clone Hero: in every runnable play path the connectome does not decide the key presses. play.py and live.py default to NullReservoir, a pass-through test double labelled 'Not a fly'; and in the prosthesis path that does build the connectome reservoir, the action is computed from the highway picture and the reservoir is stepped only to feed the side-panel visualiser, which the source says in as many words. Note also that this is the larval connectome, not the adult FlyWire/MaleCNS graph the README's framing suggests.
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.