flybrain-intransitive
Full MaleCNS LIF plays rock-paper-scissors chess with no training: captures as small-object chase, threats as looming / giant-fiber escape. Scrambled wiring collapses; a learned descending readout played worse than the biological approach/escape pair. Feature→neuron map is still hand-designed.
Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.
Dispute this rating →The evidence
No evidence, no level.LIF constants and the documented deviations from the Shiu et al. 2024 model are stated at the top of the simulation file.
https://github.com/charbelkassab/flybrain-intransitive/blob/f22c07ad64a78b9e1423c376dbc289380ef6f6fc/brain.py#L1-L30The integration loop updates voltage, conductance, adaptation and refractory state and only reads the connectome matrix.
https://github.com/charbelkassab/flybrain-intransitive/blob/f22c07ad64a78b9e1423c376dbc289380ef6f6fc/brain.py#L91-L123The connectome is built from the released MaleCNS v1.0 tables, signed by predicted transmitter.
https://github.com/charbelkassab/flybrain-intransitive/blob/f22c07ad64a78b9e1423c376dbc289380ef6f6fc/build_connectome.py#L1-L22Move choice is approach-cell minus escape-cell spike counts, with no training and a scrambled-wiring control.
https://github.com/charbelkassab/flybrain-intransitive/blob/f22c07ad64a78b9e1423c376dbc289380ef6f6fc/experiment.py#L1-L45
Whole-CNS leaky integrate-and-fire over the real MaleCNS v1.0 connectome (traced neurons, each connection signed by the presynaptic neuron's predicted transmitter), with the model and its deviations from Shiu et al. 2024 documented at the top of the file: synapse gain reduced to 0.1 mV per contact, spike-frequency adaptation added, monoamines treated as slow modulators, KC-KC contacts dropped, 1 ms step. Membrane voltage, synaptic conductance and an adaptation variable evolve each step - a documented neuron model, so L1. Not L2: nothing writes W. The matrix is loaded, filtered once by stabilise(), then only read; the move choice is a fixed readout of approach-cell minus escape-cell spike counts with no training at all, which is the project's own headline claim, and the scrambled-wiring control is included.
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.