AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

What Dense Graph Do You Need for Self-Attention?

Yuxin Wang, Chu-Tak Lee, Qipeng Guo, Zhangyue Yin, Yunhua Zhou, Xuanjing Huang, Xipeng Qiu

ICML 2022 · Architectures and efficiency › Attention and transformers · approach: Theory of self-attention dynamics · 0 citations (OpenAlex)

Read the paper

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
replace complete attention with hypercube sparse attentionTransformer · LRA averageaccuracy: 60.67 -> 62.07Improved
replace complete attention with hypercube sparse attentionTransformer · LRA at length 4096training speedup: 15.8×Improved
add random attention to Longformer-style ring lattice and global attention (ablation)Longformer-style attention · LRA averageaccuracy: 60.00 -> 59.99No clear effect
increase block size from 16 to 32 (ablation)Hypercube · LRA Retrievalaccuracy: 81.16 -> 80.74Hurt
increase block size from 16 to 32 (ablation)Hypercube · LRA Imageaccuracy: 53.79 -> 51.98Hurt
increase block size from 16 to 64 (ablation)Hypercube · LRA Retrievalaccuracy: 81.16 -> 80.75Hurt
increase block size from 16 to 64 (ablation)Hypercube · LRA Imageaccuracy: 53.79 -> 50.75Hurt
replace BERT128 with CubeBERT128BERT · Wikitext103 finetuningMLM loss: 1.18 -> 1.05Improved
use hypercube sparse attention instead of standard attentionBERT128 · GLUE averageaverage score: 82.0 -> 82.3Improved
use hypercube sparse attention instead of standard attentionBERT · GLUE averageaverage score: 83.0 -> 82.3Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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