Normalization Techniques in Training DNNs: Methodology, Analysis and Application

TL;DR

This paper reviews the evolution, methodology, and applications of normalization techniques in training deep neural networks, providing a unified framework and insights for designing new methods and understanding their role in optimization.

Contribution

It offers a comprehensive taxonomy and analysis of normalization methods, decomposing their components and discussing their applications across various tasks.

Findings

Unified view of normalization techniques from an optimization perspective

Decomposition of normalization methods into three core components

Guidance for designing new normalization approaches

Abstract

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.