5 / 10 Regularization & Loss Techniques

How does Group Normalization differ from Batch Normalization?

For a convolutional activation tensor, which description best distinguishes Group Normalization from Batch Normalization?

A

Group Normalization normalizes only the gradient tensor rather than forward activations

B

Group Normalization estimates one global mean and variance for the complete training dataset

C

Group Normalization normalizes each channel using statistics gathered across all samples in the mini-batch

D

Group Normalization partitions channels into groups and computes statistics within each sample, avoiding dependence on batch size