AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Dataset Quantization

Daquan Zhou, Kai Wang, Jianyang Gu, Xiangyu Peng, Dongze Lian, Yifan Zhang, Yang You, Jiashi Feng

ICCV 2023 · Architectures and efficiency › Quantization · approach: Post-training weight quantization · 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
replace dataset distillation (DM) with dataset quantization for cross-architecture trainingCIFAR-10 compressed dataset · CIFAR-10 (rho=10%)Avg top-1 accuracy: 39.5 -> 69.4Improved
split dataset into recursively selected bins instead of one-time coreset selection (N=1) (ablation)DQ · CIFAR-10accuracyImproved
use too many bins (N=20) (ablation)DQ · CIFAR-10accuracyHurt
drop patches by importance score instead of randomly (ablation)DQ · CIFAR-10 (rho=1%)accuracy: 41.5 -> 42.3Improved
increase patch drop ratio (ablation)DQ · CIFAR-10accuracyMixed
use DQ at 60% data keep ratio for ImageNet pretrainingResNet-50 on COCO · COCO detectionmAP: 39.2 -> 39.0No clear effect
instruction-tune LLaMA-7B on 20% DQ-selected Alpaca data instead of fullLLaMA-7B · Alpaca instruction tuningAvg of BBH, DROP, MMLU, Human-Eval: 27.7 -> 27.1No clear effect
compress with DQ instead of DM in terms of GPU hoursCIFAR-10 compression · Compression costGPU hours: 91 -> 1Improved
DQ vs random selection at 20-40% keep ratio on downstream tasksDQ pretraining subset · COCO and ADE20KmAP / mIoUImproved

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