AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Outlier Suppression: Pushing the Limit of Low-bit Transformer Language Models

Xiuying Wei, Yunchen Zhang, Xiangguo Zhang, Ruihao Gong, Shanghang Zhang, Qi Zhang, Fengwei Yu, Xianglong Liu

NeurIPS 2022 · Architectures and efficiency › Quantization · approach: Post-training quantization of LLMs · 28 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 Gamma Migration to MinMax calibration (ablation)RoBERTa 6-bit PTQ · QNLIaccuracy: 62.13 -> 78.56Improved
add Gamma Migration to MinMax calibration (ablation)RoBERTa 6-bit PTQ · CoLAMatthews corr.: 0.0 -> 0.0No clear effect
add Token-Wise Clipping (coarse stage) to baseline (ablation)RoBERTa 6-bit PTQ · QNLIaccuracy: 62.13 -> 79.46Improved
add fine-grained stage after coarse token-wise clipping (ablation)Token-Wise Clipping · SST-2accuracy: 91.63 -> 91.63No clear effect
add fine-grained stage after coarse token-wise clipping (ablation)Token-Wise Clipping · CoLAMatthews corr.: 34.95 -> 37.64Improved
add Gamma Migration to Token-Wise Clipping (ablation)RoBERTa 6-bit PTQ · CoLAMatthews corr.: 37.64 -> 46.35Improved
add Gamma Migration to Token-Wise Clipping (ablation)RoBERTa 6-bit PTQ · QNLIaccuracy: 79.66 -> 86.82Improved
use outlier suppression framework vs best baseline calibrationBERT 6-6-6 PTQ · GLUE averagescore: 74.49 -> 81.19Improved
use outlier suppression framework vs best baseline calibrationBART 6-6-6 PTQ · GLUE averagescore: 67.31 -> 79.10Improved
use outlier suppression framework with 4-4-4 QAT vs LSQ+BERT QAT · GLUE averagescore: 71.49 -> 81.13Improved
use framework with KD at 2-2-4 vs LSQ+(+KD)BERT QAT · GLUE averagescore: 66.63 -> 80.56Improved
use framework at 8-bit on SQuAD v1.1 BERT vs OMSEBERT 8-8-8 PTQ · SQuAD v1.1F1/EM: 87.90/80.16 -> 87.60/79.80Hurt

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