AI Paper Atlas

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

Generalized DataWeighting via Class-Level Gradient Manipulation

Can Chen, Shuhao Zheng, Xi Chen, Erqun Dong, Xue Liu, Hao Liu, Dejing Dou

NeurIPS 2021 · Label-efficient and robust learning › Noisy labels · approach: Loss correction and noise robustness theory

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 class-level gradient weighting (GDW) in place of instance-level weightingMWNet · CIFAR10 uniform noise 60%test accuracy (%): 81.14 -> 84.11Improved
Add class-level gradient weighting (GDW) in place of instance-level weightingMWNet · CIFAR100 uniform noise 60%test accuracy (%): 50.21 -> 53.33Improved
Add class-level gradient weighting (GDW) in place of instance-level weightingMWNet · CIFAR10 uniform noise 0%test accuracy (%): 92.95 -> 92.94No clear effect
Use GDW instead of learned label methods Soft-label/Gen-labelMeta-learning label correction · CIFAR10 uniform noise 40%test accuracy (%): 86.52 -> 88.14Improved
Use GDW instead of GLC under 40% flip noiseGLC · CIFAR10 flip noise 40%test accuracy (%): 89.74 -> 87.70Hurt
Add GDW on long-tailed dataMWNet · Long-tailed CIFAR10 imbalance 0.01test accuracy (%): 62.70 -> 71.31Improved
Fine-tune the base model on the balanced meta setBaseModel · Long-tailed CIFAR10 imbalance 0.01test accuracy (%): 69.77 -> 59.76Hurt
Fine-tune the base model on the meta setBaseModel · CIFAR10 uniform noise 60%test accuracy (%): 77.92 -> 78.41No clear effect
Remove the zero-mean constraint on class-level weights (ablation)GDW · CIFAR10 uniform noise 40%training stabilityHurt
Add GDW on a pretrained ResNet-18 on real-world noisy labelsMWNet · Clothing1Mtest accuracy (%): 68.46 -> 69.39Improved
L2RW with pretrained model instead of training from scratchL2RW · Clothing1Mtest accuracy (%): 68.80Mixed
Add GDW on a dataset with many classesMWNet · Places-LTtest accuracy (%): 18.08 -> 19.17Improved

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