AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Accelerating Training of Transformer-Based Language Models with Progressive Layer Dropping

Minjia Zhang, Yuxiong He

NeurIPS 2020 · Architectures and efficiency › Attention and transformers · approach: Efficient language transformers · 35 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 the depth-dimension schedule to temporal layer dropping (ablation)BERT + PreLN + Large lr + TD only · GLUEscore: 82.7 -> 83.2Improved
use temporal and depth-dimension schedules instead of fixed-ratio random layer dropping (ablation)BERT · GLUEscore: 82.7 -> 83.2Improved
use fixed-ratio random layer dropping instead of directly training a shallow model (ablation)BERT · GLUEscore: 81.8 -> 82.7Improved
directly train a 9-layer model without layer sampling (ablation)BERT · GLUEscore: 82.6 -> 81.8Hurt
use the full PLD schedule rather than PreLN with the large learning rateBERT · GLUEscore: 82.6 -> 83.2Improved
use the exponential temporal schedule rather than a fixed constant temporal schedule (ablation)BERT layer dropping · GLUEscore: 82.7 -> 82.7No clear effect
remove the temporal schedule (ablation)BERT layer dropping · half-precision trainingtraining stabilityHurt
use PLD instead of PreLN with the same learning rate as Post-LNBERT · GLUEscore: 80.2 -> 83.2Improved

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