seniorCNN
What is feature map normalization and why is it important in deep CNNs?
Updated May 15, 2026
Short answer
Feature map normalization stabilizes activations across layers, improving training stability and convergence speed.
Deep explanation
In deep CNNs, feature distributions can shift during training, causing unstable gradients. Normalization techniques like BatchNorm, LayerNorm, and GroupNorm standardize activations to zero mean and unit variance. This reduces internal covariate shift, enables higher learning rates, and improves generalization. It also acts as a mild regularizer.
Real-world example
Used in ResNet to stabilize training of very deep networks.
Common mistakes
- Applying normalization inconsistently between training and inference.
Follow-up questions
- What is internal covariate shift?
- How does normalization improve gradients?