deep-learningarchitecturestraining
Normalization layers
BatchNorm, LayerNorm, RMSNorm, GroupNorm — reparametrisations that stabilise deep network training.
Уровни глубины
L0Intro~1ч
Knows normalisation helps training converge.
L1Basics~6ч
Derives BatchNorm forward/backward; knows the "train vs eval" running-stats trick.
L2Working~12ч
Picks BN vs LN vs GN vs RMSNorm by architecture; handles small batches, distributed training.
L3Advanced~20ч
Understands internal-covariate-shift debate; analyses loss landscape smoothing.
L4Research~40ч
Normalisation-free networks, scaling laws, pre- vs post-norm stability.