AI Paper Atlas

Research map / Label-efficient and robust learning / Noisy labels

NoiseRank: Unsupervised Label Noise Reduction with Dependence Models

Karishma Sharma, Pinar Donmez, Enming Luo, Yan Liu, I. Zeki Yalniz

ECCV 2020 · Label-efficient and robust learning › Noisy labels · approach: Instance-dependent noise modeling · 36 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
remove top 1.8% ranked noisy instances from machine-labeled pre-training dataResNet-50 pre-trained on YFCC100m labels · ImageNet-1ktop-1 accuracy (%): 79.06 -> 79.34Improved
remove NoiseRank-ranked instances instead of the same number of random instances (ablation)ResNet-50 pre-trained on YFCC100m labels · ImageNet-1ktop-1 accuracy (%): 78.96 -> 79.34Improved
remove random 1.8% of instances instead of none (ablation)ResNet-50 pre-trained on YFCC100m labels · ImageNet-1ktop-1 accuracy (%): 79.06 -> 78.96Hurt
remove more top-ranked instances (0.6% to 1.2%) (ablation)NoiseRank pre-training filter · ImageNet-1ktop-1 accuracy (%): 79.13 -> 79.12No clear effect
fine-tune on NoiseRank-cleaned train set instead of noisy train setResNet-50 ImageNet pretrain · Clothing-1M classificationtop-1 accuracy (%): 68.94 -> 73.77Improved
initialize from the noisy-train model before fine-tuning on cleaned set (ablation)NoiseRank · Clothing-1M classificationtop-1 accuracy (%): 73.77 -> 73.82No clear effect
fine-tune on NoiseRank-cleaned train set instead of noisy train setResNet-50 ImageNet pretrain · Food-101N classificationtop-1 accuracy (%): 81.44 -> 85.20Improved
initialize from the noisy-train model before fine-tuning on cleaned set (ablation)NoiseRank · Food-101N classificationtop-1 accuracy (%): 85.20 -> 85.78Improved
add an additional clean 50k set for fine-tuningResNet-50 on noisy train · Clothing-1M classificationtop-1 accuracy (%): 68.94 -> 79.43Improved
add the clean 50k set on top of NoiseRank cleaned trainingNoiseRank · Clothing-1M classificationtop-1 accuracy (%): 73.82 -> 79.57Improved
iterate NoiseRank (fine-tune on denoised data and re-run detection) (ablation)NoiseRank · Food-101N label noise detectionaverage error rate (%): 24.02 -> 18.43Improved
iterate NoiseRank (fine-tune on denoised data and re-run detection) (ablation)NoiseRank · Clothing-1M label noise detectionaverage error rate (%): 23.54 -> 22.81Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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