Methods › General › Normalization › LayerScale

LayerScale

12 papers tagged archive 2025-07-28

Introduced by Hugo Touvron et al. in Going deeper with Image Transformers

archive 2025-07-28 Description, source and code snippet are the archive's method entry.

LayerScale is a method used for vision transformer architectures to help improve training dynamics. It adds a learnable diagonal matrix on output of each residual block, initialized close to (but not at) 0. Adding this simple layer after each residual block improves the training dynamic, allowing for the training of deeper high-capacity image transformers that benefit from depth.

Specifically, LayerScale is a per-channel multiplication of the vector produced by each residual block, as opposed to a single scalar, see Figure (d). The objective is to group the updates of the weights associated with the same output channel. Formally, LayerScale is a multiplication by a diagonal matrix on output of each residual block. In other words:

xₗ^' =xₗ+diag(λ_(l, 1), …, λ_(l, d)) ×SA(η(xₗ))

xₗ₊₁ =xₗ^'+diag(λ_(l, 1)^', …, λ_(l, d)^') ×FFN(η(xₗ^'))

where the parameters λ_(l, i) and λ_(l, i)^' are learnable weights. The diagonal values are all initialized to a fixed small value ε: we set it to ε=0.1 until depth 18 , ε=10⁻⁵ for depth 24 and ε=10⁻⁶ for deeper networks.

This formula is akin to other normalization strategies ActNorm or LayerNorm but executed on output of the residual block. Yet LayerScale seeks a different effect: ActNorm is a data-dependent initialization that calibrates activations so that they have zero-mean and unit variance, like BatchNorm. In contrast, in LayerScale, we initialize the diagonal with small values so that the initial contribution of the residual branches to the function implemented by the transformer is small. In that respect the motivation is therefore closer to that of ReZero, SkipInit, Fixup and T-Fixup: to train closer to the identity function and let the network integrate the additional parameters progressively during the training. LayerScale offers more diversity in the optimization than just adjusting the whole layer by a single learnable scalar as in ReZero/SkipInit, Fixup and T-Fixup.

PaperSource

Papers archive 2025-07-28

12 shown of 12, newest first. Repository counts are the archive's code-links table. A Syntology line states what Syntology ran from that paper's harvested code; it is per sample and not a correctness claim.

Tasks archive 2025-07-28

19 tasks the archive attaches to papers tagged with this method, by distinct papers. A task without a page in the catalog is plain text.

TaskPapers
Image Classification6
image-classification5
Semantic Segmentation3
Transfer Learning3
Classification2
Data Augmentation2
Fine-Grained Image Classification2
Object Detection2
Domain Generalization1
General Classification1
Image Segmentation1
Knowledge Distillation1
Machine Translation1
Mixture-of-Experts1
Real-Time Object Detection1
Self-Supervised Image Classification1
Self-Supervised Learning1
Speaker Verification1
Translation1

Usage over time archive 2025-07-28

Papers per year tagged with LayerScale: 2021 to 2025, peak 5 5 0 2021: 5 papers 2021 2022: 4 papers 2022 2023: 1 paper 2023 2024: 1 paper 2024 2025: 1 paper 2025
Papers per year the archive tags with this method, by the paper's archive date (12 dated). Bars are counts, not a trend claim.

Components: the archive holds no method-to-method composition, so PwC's Components table cannot be rebuilt; the Papers list carries no Results column for the same reason (the archive does not join its leaderboard rows to method tags).

Categories archive 2025-07-28

NormalizationRegularization

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections