AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

FMMformer: Efficient and Flexible Transformer via Decomposed Near-field and Far-field Attention

Tan Minh Nguyen, Vai Suliafu, Stanley Osher, Long Chen, Bao Wang

NeurIPS 2021 · Architectures and efficiency › Attention and transformers · approach: Sparse efficient transformers for long context · 5 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add Band5 near-field attention to linear transformer (1 kernel) (ablation)Linear transformer · LRAaverage test accuracy: 54.67 -> 59.51Improved
add second feature map (2 kernels) to FMMformer (ablation)FMMformer + Band5 · LRAaverage test accuracy: 59.51 -> 60.74Improved
add low-rank far-field attention to banded attention (ablation)Band5 · LRAaverage test accuracy: 57.73 -> 59.51Improved
add near-field Band5 to linear transformer on ListOps (ablation)Linear transformer · LRA ListOpstest accuracy: 18.30 -> 33.22Improved
add near-field Band5 to linear transformer on Pathfinder (ablation)Linear transformer · LRA Pathfindertest accuracy: 71.17 -> 71.29No clear effect
add linear far-field to Band5 on Retrieval (ablation)Band5 · LRA Retrievaltest accuracy: 79.41 -> 81.50Improved
add Band5 near-field to linear transformer (ablation)Linear transformer · WikiText-103test perplexity: 38.40 -> 37.29Improved
increase bandwidth from 5 to 20 (ablation)FMMformer 1-kernel · WikiText-103test perplexity: 37.29 -> 36.43Improved
use 2 kernels instead of 1 with Band20 (ablation)FMMformer · WikiText-103test perplexity: 36.43 -> 36.11Improved
replace linear far-field with fast weight update (2 kernels) (ablation)FMMformer + Band20 · WikiText-103test perplexity: 36.11 -> 34.71Improved
replace linear far-field with fast weight update (1 kernel) (ablation)FMMformer + Band20 · WikiText-103test perplexity: 36.43 -> 35.47Improved
add far-field low-rank attention to banded Band20 (ablation)Band20 · WikiText-103test perplexity: 39.19 -> 36.43Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy