@article{huang_normalization_2023,
  title = {Normalization {Techniques} in {Training} {DNNs}: {Methodology}, {Analysis} and {Application}},
  volume = {45},
  issn = {1939-3539},
  shorttitle = {Normalization {Techniques} in {Training} {DNNs}},
  doi = {10.1109/TPAMI.2023.3250241},
  number = {8},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  author = {Huang, Lei and Qin, Jie and Zhou, Yi and Zhu, Fan and Liu, Li and Shao, Ling},
  month = aug,
  year = {2023},
  note = {Conference Name: IEEE Transactions on Pattern Analysis and Machine Intelligence},
  keywords = {Batch normalization, Biological neural networks, Covariance matrices, Decorrelation, deep neural networks, image classification, Optimization, survey, Task analysis, Tensors, Training, weight normalization},
  pages = {10173--10196},
  arxiv = {arXiv:2009.12836 [cs, stat]},
  arxivdoi = {https://doi.org/10.48550/arXiv.2009.12836},
  freepdf = {https://arxiv.org/pdf/2009.12836.pdf},
  tldr = {A review and commentary of normalization methods in DNNs. I skimmed this one.  Good for definitions of all of the normalization terms and especially Figure 1.}
}

Normalization techniques are essential for accelerating the training and improving the generalization of deep neural networks (DNNs), and have successfully been used in various applications. This paper reviews and comments on the past, present and future of normalization methods in the context of DNN training. We provide a unified picture of the main motivation behind different approaches from the perspective of optimization, and present a taxonomy for understanding the similarities and differences between them. Specifically, we decompose the pipeline of the most representative normalizing activation methods into three components: the normalization area partitioning, normalization operation and normalization representation recovery. In doing so, we provide insight for designing new normalization technique. Finally, we discuss the current progress in understanding normalization methods, and provide a comprehensive review of the applications of normalization for particular tasks, in which it can effectively solve the key issues.

arXiv: http://doi.org/https://doi.org/10.48550/arXiv.2009.12836
Publisher: http://doi.org/10.1109/TPAMI.2023.3250241
pdf: https://arxiv.org/pdf/2009.12836.pdf