Group Normalization
Normalizing a fixed number of channel groups per example, independent of batch size, which is why it replaces batch normalization on small-batch vision models.
Split a convolutional layer's channels into fixed-size groups, then normalize each group's statistics within one example — no batch dimension involved, and no sequence-length assumption either. That places it between batch normalization, which needs a large batch, and layer normalization, which normalizes every channel together.
It exists for the exact spot where batch normalization degrades: object detection and segmentation models that only fit two or four high-resolution images per GPU, where batch statistics computed from so few examples are too noisy to trust.