MaleCNS on Apple MPS
Experimental PyTorch MPS simulator for Apple Silicon with verified data preparation, named-cell stimulation, and 3D activity reports.
Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.
Dispute this rating →The evidence
No evidence, no level.The neuron model is an explicit discrete LIF with documented decay, threshold and reset.
https://github.com/seohyunjun/mps-malecns-model/blob/1ae94a0698b48bde55199dac2f5dbcf86b0246d7/malecns/model.py#L16-L26Connectome weights are a registered buffer read during the step; the voltage update never writes to them.
https://github.com/seohyunjun/mps-malecns-model/blob/1ae94a0698b48bde55199dac2f5dbcf86b0246d7/malecns/model.py#L60-L78The RL trainer optimizes only the ActorCritic readout parameters with Adam and A2C loss.
https://github.com/seohyunjun/mps-malecns-model/blob/1ae94a0698b48bde55199dac2f5dbcf86b0246d7/malecns/rl.py#L106-L142The authors state the connectome is a frozen encoder: the 25.58M original connection weights are not modified and the LIF spike function is not backpropagated through.
https://github.com/seohyunjun/mps-malecns-model/blob/1ae94a0698b48bde55199dac2f5dbcf86b0246d7/README.md#L137
A PyTorch MPS simulator of MaleCNS v1.0 (166,700 neurons, 25.6M directed connections) with an explicit discrete LIF update - v[t+1] = exp(-dt/tau) v[t] + gain*W*spike[t] + drive, threshold at 1, reset to zero - and documented dt_ms/tau_ms defaults, which earns L1. The repository also ships an A2C reinforcement-learning navigation task, and it is exactly the trained-readout pattern the scale exists to separate out, not L2: the connectome weights live in a registered buffer (not an nn.Parameter), the Adam optimizer is constructed over policy.parameters() only, and the README's 'implementation scope' paragraph states outright that MaleCNS is used as a fixed activity encoder, that the 25.58M original connection weights are not changed and that the LIF spike function is not backpropagated through. The README separately lists plasticity and learning among the things the model does not have. No synapse inside the fly circuit changes at any point.
Record
Nearby
All projects →Research code for the Shiu et al. connectome-based leaky integrate-and-fire model, including activation/silencing experiments, notebooks, and FlyWire data configuration.
FlyWire whole-brain LIF implementation based on Shiu et al., with multiple simulation backends and benchmarking tools. This repository supplies the neural model; it is not a complete embodied demo package.
Ratings on thelearningfly.com are proposals, not verdicts. Every one of them is arguable in public.