How Does Batch Normalization Help Optimization?

TL;DR

This paper investigates why Batch Normalization improves neural network training, revealing that its main benefit is smoothing the optimization landscape rather than controlling input distribution shifts.

Contribution

The study challenges the common belief about internal covariate shift, showing that BatchNorm's effectiveness mainly comes from making the optimization landscape smoother.

Findings

BatchNorm does not primarily stabilize input distributions.

It significantly smooths the optimization landscape.

Smoother landscapes lead to more stable and faster training.

Abstract

Batch Normalization (BatchNorm) is a widely adopted technique that enables faster and more stable training of deep neural networks (DNNs). Despite its pervasiveness, the exact reasons for BatchNorm's effectiveness are still poorly understood. The popular belief is that this effectiveness stems from controlling the change of the layers' input distributions during training to reduce the so-called "internal covariate shift". In this work, we demonstrate that such distributional stability of layer inputs has little to do with the success of BatchNorm. Instead, we uncover a more fundamental impact of BatchNorm on the training process: it makes the optimization landscape significantly smoother. This smoothness induces a more predictive and stable behavior of the gradients, allowing for faster training.