Labs · Inside the block

Layer normalisation

Centre and scale one row to zero mean and unit variance, then let γ and β undo what training wants undone.

Try this ▶ Move the input scale to 10 and notice the blue bars spread widely — then check "Output std" below, which should still be close to 1.0 because LayerNorm rescales everything.

Computing…

Layer normalisation — Labs — TransformerLab