Transformers without Normalization

TL;DR

This paper introduces Dynamic Tanh (DyT), a simple element-wise operation that replaces normalization layers in Transformers, achieving comparable or better performance across various tasks without the need for normalization.

Contribution

The work demonstrates that normalization layers are not essential in Transformers by replacing them with DyT, a simple tanh-based operation inspired by layer normalization's input-output behavior.

Findings

Transformers with DyT match or outperform normalized models.

DyT requires minimal hyperparameter tuning.

Applicable across diverse tasks and modalities.

Abstract

Normalization layers are ubiquitous in modern neural networks and have long been considered essential. This work demonstrates that Transformers without normalization can achieve the same or better performance using a remarkably simple technique. We introduce Dynamic Tanh (DyT), an element-wise operation $DyT($x$) = \tanh(\alpha $x$)$, as a drop-in replacement for normalization layers in Transformers. DyT is inspired by the observation that layer normalization in Transformers often produces tanh-like, $S$-shaped input-output mappings. By incorporating DyT, Transformers without normalization can match or exceed the performance of their normalized counterparts, mostly without hyperparameter tuning. We validate the effectiveness of Transformers with DyT across diverse settings, ranging from recognition to generation, supervised to self-supervised learning, and computer vision to language models. These findings challenge the conventional understanding that normalization layers are indispensable in modern neural networks, and offer new insights into their role in deep networks.