그룹 정규화 퀴즈

퀴즈를 풀면 정답과 설명이 나타납니다.

1. Across what dimensions does Group Normalization compute statistics?
   • Only across examples in the batch
   • Across all model layers
   • Across the training and validation sets together
   • Spatial dimensions and channels within a group for each sample
2. Why can GroupNorm be preferable to BatchNorm for very small batches?
   • It uses the batch mean more accurately with one sample
   • Its statistics are independent of batch size
   • It removes all learned scale parameters
   • It requires a larger batch
3. How does GroupNorm divide feature channels?
   • Into groups, then normalizes each group’s values within each example
   • Into classes across the dataset
   • Into separate networks for each batch member
   • Into groups based on labels only
4. What does GroupNorm’s independence from batch size mean?
   • Its output is identical for all batch sizes and all models
   • It eliminates the need for training batches
   • Its per-example normalization does not use other examples’ statistics
   • It guarantees the same accuracy on every task
5. Which GroupNorm setting corresponds to one group containing all channels?
   • 배치 정규화
   • 레이어 정규화
   • Instance Normalization
   • Weight Normalization
6. Which GroupNorm setting corresponds to one channel per group?
   • 레이어 정규화
   • 배치 정규화
   • RMS Normalization
   • Instance Normalization
7. What does each channel’s learned scale and shift do after normalization?
   • Sets the group membership from labels
   • Computes batch statistics
   • Allows the network to adjust the normalized feature values
   • Removes the normalized activations
8. What is a careful interpretation of GroupNorm’s reported results?
   • It performed well in tested small-batch vision settings; task-specific evaluation remains needed
   • It replaces every normalization method in all networks
   • It guarantees identical results for every batch size and task
   • It removes the need to tune group count