Understanding the Disharmony between Weight Normalization Family and Weight Decay: $\epsilon-$shifted $L_2$ Regularizer

TL;DR

This paper investigates the interaction between weight normalization and weight decay, revealing theoretical insights and proposing an $oldsymbol{ ext{ extit{ extepsilon}}}$-shifted $L_2$ regularizer to improve training stability and performance.

Contribution

It provides a theoretical analysis of weight decay in the context of weight normalization and introduces a novel $ ext{ extit{ extepsilon}}$-shifted $L_2$ regularizer to address related training issues.

Findings

Weight decay only affects the effective learning rate, not generalization.

Introducing weight decay with weight normalization can cause missing global minima and instability.

The $ ext{ extit{ extepsilon}}$-shifted $L_2$ regularizer guarantees global minimum existence and improves training stability.

Abstract

The merits of fast convergence and potentially better performance of the weight normalization family have drawn increasing attention in recent years. These methods use standardization or normalization that changes the weight $\boldsymbol{W}$ to $\boldsymbol{W}'$, which makes $\boldsymbol{W}'$ independent to the magnitude of $\boldsymbol{W}$. Surprisingly, $\boldsymbol{W}$ must be decayed during gradient descent, otherwise we will observe a severe under-fitting problem, which is very counter-intuitive since weight decay is widely known to prevent deep networks from over-fitting. In this paper, we \emph{theoretically} prove that the weight decay term $\frac{1}{2}\lambda||{\boldsymbol{W}}||^2$ merely modulates the effective learning rate for improving objective optimization, and has no influence on generalization when the weight normalization family is compositely employed. Furthermore, we also expose several critical problems when introducing weight decay term to weight normalization family, including the missing of global minimum and training instability. To address these problems, we propose an $\epsilon-$shifted $L_2$ regularizer, which shifts the $L_2$ objective by a positive constant $\epsilon$. Such a simple operation can theoretically guarantee the existence of global minimum, while preventing the network weights from being too small and thus avoiding gradient float overflow. It significantly improves the training stability and can achieve slightly better performance in our practice. The effectiveness of $\epsilon-$shifted $L_2$ regularizer is comprehensively validated on the ImageNet, CIFAR-100, and COCO datasets. Our codes and pretrained models will be released in https://github.com/implus/PytorchInsight.