AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Redesigning the Transformer Architecture with Insights from Multi-particle Dynamical Systems

Subhabrata Dutta, Tanya Gautam, Soumen Chakrabarti, Tanmoy Chakraborty

NeurIPS 2021 · Architectures and efficiency › Attention and transformers · approach: Attention head and block redesign · 4 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
replace Transformer with TransEvolve-fullFF-2Transformer BASE · WMT 2014 En-FrBLEU: 38.1 -> 39.5Improved
replace Transformer with TransEvolve-fullFF-2Transformer BASE · WMT 2013 En-DeBLEU: 25.8 -> 26.2Improved
replace Transformer with TransEvolve-fullFF-2Transformer BASE · WMT 2014 En-DeBLEU: 27.3 -> 27.2No clear effect
replace feed-forward layers with random rotation matrices (randomFF vs fullFF) (ablation)TransEvolve-1 · WMT 2014 En-FrBLEU: 38.0 -> 32.6Hurt
replace feed-forward layers with random rotation matrices (randomFF vs fullFF) (ablation)TransEvolve-2 · WMT 2014 En-DeBLEU: 27.2 -> 23.8Hurt
use 2 blocks of 3-layer encoders instead of one 6-layer block (ablation)TransEvolve-fullFF · WMT 2014 En-FrBLEU: 38.0 -> 39.5Improved
use randomFF instead of fullFF (ablation)TransEvolve-1 · ListOpsaccuracy (%): 42.2 -> 43.2Improved
use 2 blocks of 3-layer encoders instead of one 6-layer block (ablation)TransEvolve-randomFF · ListOpsaccuracy (%): 43.2 -> 39.1Hurt
use randomFF instead of fullFF (ablation)TransEvolve-2 · char-level IMDBaccuracy (%): 65.6 -> 66.1Improved
use randomFF instead of fullFF (ablation)TransEvolve-1 · AGnewsaccuracy (%): 91.1 -> 90.6Hurt
use randomFF instead of fullFF (ablation)TransEvolve-2 · IMDBaccuracy (%): 87.6 -> 87.5No clear effect
replace Transformer with TransEvolve-fullFF-1Transformer · AGnewsaccuracy (%): 88.8 -> 91.1Improved

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