Spatial SNN Ablation — Training Curves

	name	cell	topology	λ	hidden	best test acc	Δ vs dense	ev/n/s	K̄ rec	% of dense	spec gap	alg conn	acc / param	score (vs same-N)	lr	p₀	batch	T steps	wd	seed	params (effective)	epochs done	s / epoch	ETA / done	wall

topology overview — stochastic patterns

For each topology: left = adjacency matrix (white pixel = connection); middle = spatial layout, connections from one centred neuron drawn in topology colour; right = fan-in distribution across neurons. N=128, seed=0; orig rows at λ̃=0.55/p₀=0.25; LOC rows at periodic torus + sharper kernel.

topology overview — regular (translation-invariant) patterns

Three candidate regular patterns: same offset set for every neuron on the [0,1)³ torus, so K is exactly the same per neuron and the kernel is tensor-core friendly. 4-panel layout: adjacency matrix, 3D connectivity from one centred neuron, fan-in histogram (single delta), and the offset kernel itself (the relative-position pattern that's reused for every neuron).