batch normalization layer

**Normalization Techniques** are the **layer-level operations that standardize activations within a neural network during training — reducing internal covariate shift, stabilizing gradient flow, and enabling higher learning rates that accelerate convergence, with different variants (Batch, Layer, Group, RMS normalization) suited to different architectures, batch sizes, and deployment scenarios**. **Why Normalization Is Necessary** As data flows through a deep network, the distribution of activations at each layer shifts with every parameter update (internal covariate shift). Without normalization, deeper layers must constantly adapt to changing input distributions, slowing training and requiring careful initialization and low learning rates. Normalization fixes the input distribution at each layer, decoupling layers and allowing independent, faster learning. **Batch Normalization (BatchNorm)** The original breakthrough (Ioffe & Szegedy, 2015): - **During training**: For each channel, compute mean and variance across the batch dimension and spatial dimensions (B, H, W). Normalize: x_hat = (x - μ) / √(σ² + ε). Apply learned affine transform: y = γ × x_hat + β. - **During inference**: Use running mean/variance accumulated during training (not batch statistics), making inference deterministic and independent of batch composition. - **Limitation**: Requires sufficiently large batch sizes (≥16-32) for stable statistics. Breaks down with batch size 1 (inference on single samples uses running stats, but fine-tuning is problematic). Not suitable for sequence models where the batch dimension has variable-length inputs. **Layer Normalization (LayerNorm)** Computes statistics across the feature dimension for each individual sample (not across the batch): - **Normalization axis**: All features within a single token/sample. For a Transformer with hidden dim 768, mean and variance computed over those 768 values per token. - **Advantage**: Independent of batch size — works with batch size 1 and variable-length sequences. The default normalization for Transformers (GPT, BERT, LLaMA). - **Pre-LayerNorm vs. Post-LayerNorm**: Pre-LN (normalize before attention/FFN) stabilizes training of very deep Transformers, enabling training without learning rate warmup. **Group Normalization (GroupNorm)** Divides channels into groups (typically 32) and normalizes within each group per sample. Combines BatchNorm's channel-wise normalization with LayerNorm's batch-independence. Preferred for computer vision tasks with small batch sizes (object detection, segmentation where high-resolution images limit batch size). **RMSNorm** A simplified LayerNorm that normalizes by the root mean square only (no mean subtraction): y = x / RMS(x) × γ. Removes the mean computation, reducing overhead by ~10-15%. Used in LLaMA, Gemma, and modern LLMs where the marginal speedup at scale is significant. **Impact on Training Dynamics** Normalization layers act as implicit regularizers — the noise in batch statistics (BatchNorm) or the constraint on activation scale provides a regularization effect similar to dropout. Networks with normalization typically need less dropout and less careful weight initialization. Normalization Techniques are **the critical infrastructure that makes deep network training stable and efficient** — a seemingly simple statistical operation that transformed deep learning from a fragile art requiring careful initialization into a robust engineering practice where networks of arbitrary depth train reliably.