Fruit Fly Fashion
Reproducible art experiment using MaleCNS spike vectors to control the placement, rotation, and scale of print designs.
Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.
Dispute this rating →The evidence
No evidence, no level.The pipeline imports DOOMFLY's NativeBrain, steps it 150 ms per pass on retinal input, and hashes the spike counts into the art transform parameters.
https://github.com/jtc268/fruit-fly-fashion/blob/ca76eb47423e13238cd84f390dc877ce2cf14bb1/scripts/run_atelier.py#L245-L283The pinned DOOMFLY kernel is a leaky integrate-and-fire integrator with 20 ms membrane and 5 ms synaptic time constants, -52 mV rest and a 1.8 ms delay / 2.2 ms refractory period.
https://github.com/nftechie/doomfly/blob/71ecf53d78eaffaf1a57ed7b0ccf5d458abc9f33/doom/kernel.cpp#L12-L21The connectome weight array is declared const in the kernel signature, so synaptic weights cannot change during simulation; spikes only read weight[e] into the postsynaptic conductance.
https://github.com/nftechie/doomfly/blob/71ecf53d78eaffaf1a57ed7b0ccf5d458abc9f33/doom/kernel.cpp#L7-L11The authors' own 'Honest label' section says the dynamics and retina are explicit models, not a resurrected biological fly.
https://github.com/jtc268/fruit-fly-fashion/blob/ca76eb47423e13238cd84f390dc877ce2cf14bb1/README.md#L53-L55
The art pipeline really does run the full MaleCNS graph (166,700 nodes, 25.6M edges) through a LIF simulation each pass and hash the resulting spike-count vector into rotation/scale/placement parameters. The neuron model lives in the pinned DOOMFLY submodule (commit 71ecf53) and is a documented LIF: exponential subthreshold decay with a 20 ms membrane constant, 5 ms synaptic constant, -52 mV rest/reset, -45 mV threshold, 2.2 ms refractory, 1.8 ms delay. That is L1. Not L2: the connectome weight array is passed into the native kernel as `const float* weight` - the integrator literally cannot write to it - and `run_atelier.py` resets v, g, drive and refractory state between products rather than carrying any adaptation forward. The 'learning' in the pipeline is a SHA-256 hash of the spike vector driving image transforms, which is a mapping, not plasticity. Note for reviewers: the DOOMFLY repo also ships a `doom_learning/` tree (including `conditioning.py`) that this project does not import - only `doom.native.NativeBrain` and `doom.game.retinal_samples` are used.
Record
Nearby
All projects →Ratings on thelearningfly.com are proposals, not verdicts. Every one of them is arguable in public.