fly-brain

A faithful re-implementation of the Shiu et al. whole-brain LIF model across five simulation backends (Brian2 CPU/CUDA, PyTorch, NEST GPU, GeNN, Brian2GeNN), with the same documented constants: 20 ms membrane, 5 ms alpha synapse, -52 mV rest/reset, -45 mV threshold, 2.2 ms refractory, 1.8 ms delay, w_syn 0.275. Membrane voltage is integrated explicitly each timestep, so L1. Not L2: the connectome weight matrix is loaded from the FlyWire connectivity table, cached as a sparse tensor and used read-only in the recurrent matmul; a repo-wide search over every Python file at this commit finds no `torch.optim`, no optimizer, no `loss.backward()`, no learning rate, no STDP and no dopamine. The PyTorch backend does define an ATan surrogate-gradient autograd function, which reads like training infrastructure, but nothing in this repository ever constructs an optimizer or takes a gradient step - it is a benchmarking and cross-backend-parity harness. No validation mark either: the 'ground truth comparison' in this repo compares backends against Brian2 CPU, i.e. a numerical-agreement check between simulators, not a comparison against measurements from real flies; the 91% accuracy figure quoted in the README is Shiu et al.'s result, not something reproduced here.