Kick the Fly
Kick-the-buddy toy on a live 166,700-neuron MaleCNS. Tools hit real sensory cells; sugar lights PAM dopamine and heals. Reward is the sugar pulse, not a trained game policy.
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: dopamine-gated depression scaled by KC eligibility, floored at 10% of the connectome weight, with opposite-compartment reversal and slow forgetting.
https://github.com/legendarylolo318-cloud/kick-the-fly/blob/fad484cea6173c7fd7522cc66a98dd2f9556802f/memory.py#L147-L167The changed weights are written directly into the simulator's own synapse matrices, so the LIF sim runs on them.
https://github.com/legendarylolo318-cloud/kick-the-fly/blob/fad484cea6173c7fd7522cc66a98dd2f9556802f/memory.py#L169-L171The plastic synapses are selected as the KC->MBON entries of the simulator's connectome matrix, not a separate readout layer.
https://github.com/legendarylolo318-cloud/kick-the-fly/blob/fad484cea6173c7fd7522cc66a98dd2f9556802f/memory.py#L86-L96DAN and MBON identity and the per-MBON dopamine gate come from the connectome's cell types and its DAN->MBON synapse counts.
https://github.com/legendarylolo318-cloud/kick-the-fly/blob/fad484cea6173c7fd7522cc66a98dd2f9556802f/brainpack.py#L57-L59Plasticity runs inside the live brain loop every 10 simulation steps, not in an offline training pass.
https://github.com/legendarylolo318-cloud/kick-the-fly/blob/fad484cea6173c7fd7522cc66a98dd2f9556802f/kick_the_fly.py#L399-L401The underlying simulation is leaky integrate-and-fire over the signed connectome with stated parameters (tau_m 20 ms, threshold, refractory period), which is the L1 floor this project builds on.
https://github.com/legendarylolo318-cloud/kick-the-fly/blob/fad484cea6173c7fd7522cc66a98dd2f9556802f/connectome/sim.py#L32-L47
Real dopamine-gated plasticity on the connectome's own synapses, verified to the assignment. The plastic set is the KC->MBON entries of the simulator's synapse matrix itself, located by walking the CSR index arrays; the update depresses them by rate * dopamine * eligibility * weight (floored at 10% of the connectome value, with reversal and a slow decay back toward it), and `_write_back` assigns the result straight into `sim.W_csr.data` / `sim.W_csc.data` — the same matrices the LIF step propagates through. Which DANs gate which MBON, and whether a compartment is PPL1 (punishment) or PAM (reward), is read from the connectome's own DAN->MBON synapse-count matrix, not hand-assigned. It runs live in the brain thread every 10 sim steps during play, not in an offline training script. No validation mark: the rule is grounded in cited fly work (Aso 2014, Hige 2015, Cohn 2015, Felsenberg 2018), but there is no published comparison of this model's behaviour against real fly measurements. The authors also flag which parts are game rules rather than biology — pain driving PPL1, sugar driving PAM, the tool-to-odour mapping, and the learning and forgetting rates.
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.