Fly Lab
Ableton Live music experiments driven by simulated fly motor circuits and a small trainable musical readout. The measured wiring remains fixed.
Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.
Dispute this rating →The evidence
No evidence, no level.The locomotor subgraph runs a spiking model with membrane leak, threshold, reset, refractory period and adaptation over measured MaleCNS contact counts.
https://github.com/Apolotary/fly-lab/blob/631ada7f1e074581ac986e91c7e68ad7074fedb1/src/vendor/desktop-fly/locomotor.js#L103-L128The only learning is a policy-gradient update on a small separate action-by-feature weight matrix, driven by an external phrase reward.
https://github.com/Apolotary/fly-lab/blob/631ada7f1e074581ac986e91c7e68ad7074fedb1/src/dark-learner.js#L108-L140The authors state directly that learning changes 70 musical weights and the measured fly wiring stays fixed.
https://github.com/Apolotary/fly-lab/blob/631ada7f1e074581ac986e91c7e68ad7074fedb1/README.md#L5-L11
A measured MaleCNS v1.0 leg locomotor subgraph is simulated with a genuine spiking model - leaky membrane decay, a firing threshold, a reset, a refractory period, spike-frequency adaptation and separate excitatory/inhibitory synaptic conductances with their own decay constants - which earns L1. Not L2: the thing that learns is a 70-weight contextual policy bolted on after the six motor outputs, trained by a REINFORCE-style policy gradient with an entropy bonus, and it identifies itself in its own telemetry as a 'trained-musical-readout'. The connectome weight array is written exactly once, at construction, to normalize contact counts into synaptic gains; nothing in the project ever changes a synapse afterwards. The README says this plainly and without hedging: the learning changes 70 musical weights and the measured fly wiring stays fixed.
Record
Nearby
All projects →Ratings on thelearningfly.com are proposals, not verdicts. Every one of them is arguable in public.