AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Low-Rank Bottleneck in Multi-head Attention Models

Srinadh Bhojanapalli, Chulhee Yun, Ankit Singh Rawat, Sashank Reddi, Sanjiv Kumar

ICML 2020 · Architectures and efficiency › Attention and transformers · approach: Long-context KV and memory attention · 14 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
Increase heads from 8 to 32 with the standard head-size heuristic24-layer Transformer, embedding size 1024 · SQuAD devF1: 90.89$\pm$0.15 -> 90.45$\pm$0.08Hurt
Increase heads from 8 to 32 with the standard head-size heuristic24-layer Transformer, embedding size 1024 · SQuAD devEM: 84.1$\pm$0.34 -> 83.48$\pm$0.13Hurt
Increase heads from 8 to 32 with the standard head-size heuristic24-layer Transformer, embedding size 1024 · MNLI devaccuracy: 85$\pm$0.2 -> 84.4$\pm$0.2Hurt
Fix head size at 128 while increasing heads from 8 to 32 (ablation)FixedMultiHead, embedding size 512 · SQuAD devF1: 89.6$\pm$0.17 -> 90.95$\pm$0.14Improved
Fix head size at 128 while increasing heads from 8 to 32 (ablation)FixedMultiHead, embedding size 512 · SQuAD devEM: 82.73$\pm$0.21 -> 84.4$\pm$0.29Improved
Fix head size at 128 while increasing heads from 8 to 32 (ablation)FixedMultiHead, embedding size 512 · MNLI devaccuracy: 83.5$\pm$0.2 -> 84.9$\pm$0.2Mixed
Increase head size from 32 to 256 (ablation)FixedMultiHead, embedding size 512, 8 heads · SQuAD devF1: 88.53$\pm$0.06 -> 90.33$\pm$0.23Improved
Increase head size from 32 to 256 (ablation)FixedMultiHead, embedding size 512, 8 heads · SQuAD devEM: 81.19$\pm$0.21 -> 83.36$\pm$0.48Improved
Increase head size from 32 to 256 (ablation)FixedMultiHead, embedding size 512, 8 heads · MNLI devaccuracy: 82.5$\pm$0.1 -> 83.9$\pm$0.2Improved
Set head size to input sequence length, decoupled from embedding size and head countTransformer · SQuAD and MNLI devperformanceImproved
Use embedding size 512 with fixed head size instead of the 1024-dimensional baseline24-layer Transformer · SQuAD and MNLI devperformanceImproved

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