Methods › General › Attention Modules › Multi-Head Linear Attention

Multi-Head Linear Attention

24 papers tagged archive 2025-07-28

Introduced by Sinong Wang et al. in Linformer: Self-Attention with Linear Complexity

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

Multi-Head Linear Attention is a type of linear multi-head self-attention module, proposed with the Linformer architecture. The main idea is to add two linear projection matrices Eᵢ, Fᵢ ∈ℝ^(n×k) when computing key and value. We first project the original (n ×d)-dimensional key and value layers KWᵢᴷ and VWᵢⱽ into (k×d)-dimensional projected key and value layers. We then compute a (n×k) dimensional context mapping P̅ using scaled-dot product attention:

h̅e̅a̅d̅ᵢ̅ = Attention(QW^Qᵢ, EᵢKWᵢᴷ, FᵢVWᵢⱽ)

h̅e̅a̅d̅ᵢ̅ = softmax((QW^Qᵢ(EᵢKWᵢᴷ)ᵀ)/(√(dₖ))) ·FᵢVWᵢⱽ

Finally, we compute context embeddings for each head using P̅ ·(FᵢVWᵢⱽ).

PaperSourceSee Code · tatp22/linformer-pytorch

Papers archive 2025-07-28

24 shown of 24, 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

20 shown of 48 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
Knowledge Distillation3
Language Modelling3
Classification2
GPU2
Language Modeling2
Prediction2
Survey2
Abstractive Text Summarization1
Active Learning1
Autonomous Driving1
Cloud Removal1
Data Augmentation1
Deblurring1
Decoder1
Depth Estimation1
Face Recognition1
Generative Adversarial Network1
Image Classification1
Image Generation1
Image Restoration1

Usage over time archive 2025-07-28

Papers per year tagged with Multi-Head Linear Attention: 2020 to 2025, peak 6 6 0 2020: 4 papers 2020 2021: 6 papers 2021 2022: 4 papers 2022 2023: 3 papers 2023 2024: 5 papers 2024 2025: 2 papers 2025
Papers per year the archive tags with this method, by the paper's archive date (24 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

Attention Modules

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