Batch Normalization

Step-by-step · mean · variance · normalize · scale & shift

Batch mean (μ)
Batch variance (σ²)
Normalized (x - μ)/√(σ²+ε)
Scaled & shifted (γ·norm+β)
Input values
Inference mode — During training, running mean & variance are updated via exponential moving average. At inference, the running mean and running variance replace batch statistics. This ensures consistent normalization regardless of batch size.