AI Paper Atlas

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

Regularly Truncated M-Estimators for Learning With Noisy Labels

Xiaobo Xia, Pengqian Lu, Chen Gong, Bo Han, Jun Yu, Jun Yu, Tongliang Liu

TPAMI 2023 · Label-efficient and robust learning › Noisy labels · approach: Sample selection and co-training · 23 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 regular switching back to the original M-estimator (large-loss examples periodically included) (ablation)Truncated Catoni's · CIFAR-10 Sym-30%test accuracy (%): 86.43 -> 89.39Improved
add regular switching back to the original M-estimator (ablation)Truncated Catoni's · CIFAR-10 Pair-30%test accuracy (%): 90.25 -> 90.83Improved
add regular switching back to the original M-estimator (ablation)Truncated Welsch+ · CIFAR-10 Ins-30%test accuracy (%): 88.50 -> 89.56Improved
add truncation to the original Catoni's M-estimator (ablation)Catoni's M-estimator · CIFAR-10 Sym-30%test accuracy (%): 84.95 -> 86.43Improved
add truncation to Log-sum M-estimator (ablation)Log-sum · CIFAR-10 Pair-30%test accuracy (%): 54.64 -> 88.44Improved
assign different weights to small-loss examples (M-estimator weights instead of equal CE weights) (ablation)Truncated CE · CIFAR-10 Pair-30% (Catoni's)test accuracy (%): 86.32 -> 90.25Improved
increase switching period R (ablation)RT M-estimators · synthetic MNIST and CIFAR-10test accuracy (%)Hurt
perturb truncation threshold by -20% for truncated M-estimators (ablation)Truncated M-estimators · synthetic CIFAR-10 30% noisetest accuracy (%): reduced by almost 20%Hurt
estimate epsilon and alpha adaptively via Gaussian assumption instead of fixed values (ablation)RT-Catoni's/Log-sum/Welsch+ · CIFAR-10 Sym-30%test accuracy (%): 89.05 -> 89.36 (Catoni's); 89.60 -> 89.42 (Log-sum)No clear effect
add RT-Catoni's warm-up to DivideMixDivideMix · CIFAR-10 Pair-45%test accuracy (%): 86.55 -> 94.88Improved
add RT-Welsch warm-up to DivideMixDivideMix · CIFAR-10 Ins-50%test accuracy (%): 92.90 -> 91.51Hurt
add RT-Catoni's fine-tuning to contrastive learning (MoCo v2)CL · CIFAR-10 Ins-50%test accuracy (%): 56.51 -> 75.45Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 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