Shows a batch of activation vectors (rows) being normalized per-sample across feature dimensions: compute μ and σ² for each row, normalize z=(x-μ)/√(σ²+ε), then apply learned per-feature affine parameters γ and β to produce the output y. Animation highlights one sample at a time to emphasize that statistics are not computed across the batch.
Pure Canvas2D, blocky grid cells with snapped coordinates. 4-stage loop (input → stats → normalize → affine) over ~4.2s; active sample and feature indices cycle independently. Per-sample μ/σ² bars and per-feature γ/β panel reinforce the axes of normalization vs learned affine parameters.