AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

NxMTransformer: Semi-Structured Sparsification for Natural Language Understanding via ADMM

Connor Holmes, Minjia Zhang, Yuxiong He, Bo Wu

NeurIPS 2021 · Architectures and efficiency › Pruning and efficient tuning · approach: Low-rank compression and merging · 6 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
use ADMM-based NxMTransformer instead of one-shot magnitude pruning (ASP) for 4:2 sparsityBERT-base · GLUE dev averageaverage score: 78.8 -> 80.5Improved
use NxMTransformer instead of ASP on small tasks (CoLA, STS-B, MRPC, RTE)BERT-base 4:2 · GLUE small tasksaccuracy: 51.7 -> 55.3 on CoLAImproved
use NxMTransformer instead of ASP on RTEBERT-base 4:2 · RTEaccuracy: 63.9 -> 68.6Improved
use NxMTransformer instead of ASP on large tasks (MNLI, QNLI, SST-2)BERT-base 4:2 · MNLI-maccuracy: 83.3 -> 82.3No clear effect
use NxMTransformer instead of ASP on QNLIBERT-base 4:2 · QNLIaccuracy: 90.6 -> 90.4No clear effect
use NxMTransformer instead of ASP, median over 7 random seeds on tasks with fewer than 10,000 examplesBERT-base 4:2 · small GLUE tasksaccuracy (median): +1.9 pointsImproved
use unstructured 50% ADMM sparsity instead of 4:2 semi-structured sparsityBERT-base · GLUE dev averageaverage score: 80.5 -> 81.3Improved
sparsify dense BERT-base to 4:2 with NxMTransformerBERT-base · GLUE dev averageaverage score: 81.8 -> 80.5Hurt
apply NxMTransformer 4:2 sparsity to DistilBERTDistilBERT · GLUE dev averageaverage score: 78.5 -> 76.6Hurt
use 4:2 sparse BERT-base (NxMTransformer) instead of DistilBERT at same parameter countBERT-base vs DistilBERT · GLUE dev averageaverage score: 78.5 -> 80.5Improved
increase mask similarity across ADMM iterations by tuning rho (up to about 99% similarity) (ablation)NxMTransformer · CoLA and QNLIMatthews correlation / accuracyImproved
increase mask similarity beyond about 99% (rho too large) (ablation)NxMTransformer · CoLA and QNLIaccuracyHurt

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