AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

The Emergence of Essential Sparsity in Large Pre-trained Models: The Weights that Matter

AJAY KUMAR JAISWAL, Shiwei Liu, Tianlong Chen, Zhangyang Wang

NeurIPS 2023 · Architectures and efficiency › Pruning and efficient tuning · approach: Lottery ticket discovery · 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
Prune lowest-magnitude weights one-shot before fine-tuningbert-base · GLUE and SQuAD downstream tasksfine-tuning performance: 21.50No clear effect
Prune OPT-1.3B beyond 40% sparsity vs bert-base beyond 60%OPT-1.3B · RTEfine-tuning performance: 40% vs 60%Hurt
Use self-supervised DINO pre-training instead of supervised ViTViT-base architecture · CIFAR-10, CIFAR-100, Tiny-ImageNetzero weights and robustness to pruning: 14.37Improved
Increase pre-training data volume from 60% to 100% of bookcorpus (ablation)bert-base · MNLIprunability and downstream accuracyImproved
Replace expensive LTH masks with free OMP masks within essential sparsity range (ablation)bert-base, ViT-base · multiple downstream tasksfine-tuning performance: $>96\%$No clear effect
Use SparseGPT instead of OMP for pruning (ablation)Vicuna-7B · MMLUperformance at high sparsityImproved
Apply N:M structured pruning instead of unstructured (ablation)OPT-350M · MNLI, QNLI, RTE, SST-2fine-tuning performance dropNo clear effect

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