Encyclopedia · 176 concepts

Architectures · intermediate · concept 61 of 176

Normalization Layers (BatchNorm to RMSNorm)

Layers that rescale activations so training stays stable as networks get deep. BatchNorm made deep CNNs trainable, LayerNorm made transformers possible, and RMSNorm is the leaner variant inside most current LLMs. Unglamorous, and load-bearing for everything.

Key terms

Batch normalizationLayer normalizationRMSNormInternal covariate shiftTraining stability

Where you meet it in the real world

Every deep network trained since 2015; pre-norm vs post-norm is a real LLM design decision