flyvis

The official implementation of Lappalainen et al., Nature 2024. Neuron dynamics are explicit and documented: passive point neurons with instantaneous graded-release synapses, dv/dt = (1/tau) * (-v + bias + sum_j w_ji * relu(v_j) + x_t), with a per-cell-type membrane time constant. That is a documented neuron model with state, so L1. This is the batch's hardest call, because unlike the readout-only projects the trained parameters here really are inside the fly circuit: the edge weight is sign * synapse_count * syn_strength, and syn_strength is an nn.Parameter. It still does not reach L2, for two reasons the scale calls out directly. First, what changes those weights is loss.backward() and an Adam step against an optic-flow task loss - an optimizer's gradient, not a rule grounded in fly biology; a search of the whole package for a Hebbian, STDP, dopaminergic or eligibility term returns nothing at all. Second, it is offline training that bakes in fixed weights: during a simulation or analysis run the parameters are constant, and the shipped pretrained ensembles are exactly such frozen weights. Validated is a separate question from the level, and this project earns it: the repository ships direction-tuning curves digitised from Maisak et al. 2013 recordings in live flies and correlates the model's responses against them, and the Nature paper publishes that comparison across the visual system. It is a comparison of neural activity rather than of behaviour, which the mark covers, and the distinction is worth keeping.