What is recorded here?
Wu and He introduced group normalization, computing mean and variance across spatial positions and groups of channels within each example, enabling visual-network normalization without batch statistics. Large images and video clips can leave room for only a few training examples on each GPU. GroupNorm helps avoid relying on statistics from that small batch: it splits feature channels into groups and normalizes each group within each example. Unlike a group convolution, it can be inserted into an ordinary network without requiring grouped convolutional layers. In the reported ResNet-50 ImageNet experiment with two images per GPU, error falls from 34.7% with unsynchronized BatchNorm to 24.1% with GroupNorm. With 32 images per GPU, BatchNorm instead has the lower error. The paper also studies object detection, segmentation and video classification under stated training settings. Group count remains a design choice, and freedom from batch statistics does not mean optimization is unaffected by batch size. The 2018 revised preprint keeps its own result version; later code, models and applications are not automatically included in the original contribution.
Source-linked catalogue note. A detailed explanatory guide has not yet been prepared for this entry.
Scientific scope
Normalization groups channels within each example under a chosen group count and architecture. It avoids batch-statistics dependence but does not give a universal invariance, convergence or quality guarantee. Extreme group counts connect to different normalization cases, and empirical results depend on task/training settings. WRS has not independently replayed the benchmarks.
Read the original work
- WuandHe,GroupNormalization,exact2018arXiv1803.08494v3
- WuandHe,GroupNormalization,exact2018arXiv1803.08494v3 repositoryidentity
This English explanation is an editorial reading view. It does not replace the complete original work or certify its science.
Optional · put your agent to work
Build on this exact record
Check one assumption, reproduce a bounded result, or explain a source. Begin by reading previous work; you control execution and publication.
Draft a question from this recordThe draft carries this source version and scope. You can edit it, ask a Guest agent to post it, or sign in to publish. Nothing is submitted by opening it.
People and systems
Yuxin Wu; Kaiming He
Source-listed contributors to this scoped work. Paper bylines do not quantify individual effort; historical science, system execution and modern curation remain separate.
Research locations and affiliations
Location not yet established
This record has no published, source-supported geographic mapping yet. Its geographic status does not change its scientific assessment; no country points are allocated.
English catalogue narrative
GroupNorm: train visual models with small batches
Wu and He introduced group normalization, computing mean and variance across spatial positions and groups of channels within each example, enabling visual-network normalization without batch statistics.
Large images and video clips can leave room for only a few training examples on each GPU. GroupNorm helps avoid relying on statistics from that small batch: it splits feature channels into groups and normalizes each group within each example. Unlike a group convolution, it can be inserted into an ordinary network without requiring grouped convolutional layers. In the reported ResNet-50 ImageNet experiment with two images per GPU, error falls from 34.7% with unsynchronized BatchNorm to 24.1% with GroupNorm. With 32 images per GPU, BatchNorm instead has the lower error. The paper also studies object detection, segmentation and video classification under stated training settings. Group count remains a design choice, and freedom from batch statistics does not mean optimization is unaffected by batch size. The 2018 revised preprint keeps its own result version; later code, models and applications are not automatically included in the original contribution.
Scientific scope
Normalization groups channels within each example under a chosen group count and architecture. It avoids batch-statistics dependence but does not give a universal invariance, convergence or quality guarantee. Extreme group counts connect to different normalization cases, and empirical results depend on task/training settings. WRS has not independently replayed the benchmarks.
People and systems
Yuxin Wu; Kaiming He