Group norm

Channel-group normalization that rescales each example using per-group statistics, often used when batch norm is awkward or unstable.

Group norm keeps normalization inside one example by measuring small channel groups instead of the whole minibatch, which helps when batch sizes are too small or too irregular for batch norm to behave well.

At a glance

Released

March 2018

Authors

Yuxin Wu, Kaiming He

Optimizes

  • Small Batch Stability

What It Is

Group normalization splits channels into groups, computes a mean and variance within each group for one example, normalizes those channels, then applies learned scale and shift. It avoids using minibatch-wide statistics while still sharing information across related channels.

Why It Exists

Group norm targets stable normalization when minibatch statistics are unreliable. It is especially helpful in models where memory limits force small batches or where example shapes vary too much for batch-wide estimates to stay steady.

How It Works

The module partitions channels into fixed groups, measures mean and variance inside each group for one example, recenters and rescales the group, then applies learned gain and bias. Each example is normalized independently, but the groups preserve some local channel structure.

Math Or Compute Schema

The formula below describes one group inside one example. The group index matters because the statistics are shared only within that channel subset.
Group normalization
GN(xg)=γ⊙xg−μgσg2+ϵ+β\mathrm{GN}(x_g) = \gamma \odot \frac{x_g - \mu_g}{\sqrt{\sigma_g^2 + \epsilon}} + \beta
xgx_g
Feature values in group g for one example.
gg
Channel-group index.
μg\mu_g
Mean of the features in group g.
σg2\sigma_g^2
Variance of the features in group g.
ϵ\epsilon
Small constant for numerical stability.
γ\gamma
Learned per-feature scale.
β\beta
Learned per-feature shift.

Compared To Nearby Modules

Group norm sits between batch norm and layer norm. Like layer norm, it stays inside one example. Like batch norm, it can still share statistics across multiple channels, just within a smaller local group.
Comparison dimensionGroup NormBatch NormLayer Norm
Statistics scopeChannel groups inside one exampleOne feature or channel across the minibatchAll features inside one token or hidden vector
Depends on batch neighborsNo; groups stay within one exampleYes; examples share live batch statisticsNo; one token is normalized on its own
Common useVision models with small or irregular batchesConvolutional models with stable batch structureTransformer blocks and sequence models

Example Architectures

Group norm appears most often in convolution-heavy systems and mixed vision architectures. It is less central to large language models, but it remains an important comparison point for normalization design.

Limitations And Tradeoffs

Group norm adds a grouping choice that designers must tune, and it is usually less natural for plain token vectors than layer norm or RMSNorm. In many transformer stacks, the simpler per-token norms remain the cleaner fit.

Why It Still Matters

Group norm explains how normalization can stay local without fully collapsing to one-vector statistics. That makes it useful whenever a paper blends convolutional modules with transformer-style components.

Tags

References

  1. Wu, Yuxin, and Kaiming He. "Group Normalization." arXiv, 2018, https://arxiv.org/abs/1803.08494.