Group Normalization

TL;DR

Group Normalization (GN) offers a batch size-independent normalization method that maintains accuracy across various batch sizes, outperforming Batch Normalization in small-batch scenarios and effectively transferring to downstream tasks.

Contribution

This paper introduces Group Normalization as a simple, effective alternative to Batch Normalization that is stable across different batch sizes and easily applicable to various computer vision tasks.

Findings

GN reduces error by 10.6% compared to BN at batch size 2.

GN performs comparably or better than BN at typical batch sizes.

GN outperforms BN in object detection, segmentation, and video classification tasks.

Abstract

Batch Normalization (BN) is a milestone technique in the development of deep learning, enabling various networks to train. However, normalizing along the batch dimension introduces problems --- BN's error increases rapidly when the batch size becomes smaller, caused by inaccurate batch statistics estimation. This limits BN's usage for training larger models and transferring features to computer vision tasks including detection, segmentation, and video, which require small batches constrained by memory consumption. In this paper, we present Group Normalization (GN) as a simple alternative to BN. GN divides the channels into groups and computes within each group the mean and variance for normalization. GN's computation is independent of batch sizes, and its accuracy is stable in a wide range of batch sizes. On ResNet-50 trained in ImageNet, GN has 10.6% lower error than its BN counterpart when using a batch size of 2; when using typical batch sizes, GN is comparably good with BN and outperforms other normalization variants. Moreover, GN can be naturally transferred from pre-training to fine-tuning. GN can outperform its BN-based counterparts for object detection and segmentation in COCO, and for video classification in Kinetics, showing that GN can effectively replace the powerful BN in a variety of tasks. GN can be easily implemented by a few lines of code in modern libraries.