How does Group Normalization differ from Batch Normalization?

Question: For a convolutional activation tensor, which description best distinguishes Group Normalization from Batch Normalization? A) Group Normalization estimates one global mean and variance for the complete training dataset B) Group Normalization normalizes each channel using statistics gathered across all samples in the mini-batch C) Group Normalization partitions channels into groups and computes statistics within each sample, avoiding dependence on batch size D) Group Normalization normalizes only the gradient tensor rather than forward activations Correct: C Explanation: Group Normalization divides channels into groups and calculates normalization statistics within each individual sample. It therefore avoids BatchNorm's dependence on reliable mini-batch statistics and is often useful when batches are small. Topic: advanced ML / normalization / GroupNorm