Research map / Label-efficient and robust learning / Noisy labels
Retaining Beneficial Information from Detrimental Data for Neural Network Repair
Long-Kai Huang, Peilin Zhao, Junzhou Huang, Sinno Pan
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 changed | Where | Measured | Effect |
|---|---|---|---|
| align identified data with energy-based alignment instead of deleting it (BIR vs EWC-Deletion), imprecise cause set | model treatment, LeNet · MNIST-Label (75% cause + 25% clean) | test accuracy (All, %): 87.74 -> 95.70 | Improved |
| use imprecise rather than precise cause set for BIR | BIR, LeNet · MNIST-Label | test accuracy (All, %): 95.64 -> 95.70 | No clear effect |
| use imprecise rather than precise cause set for EWC-Deletion | EWC-Deletion, LeNet · MNIST-Label | test accuracy (All, %): 95.10 -> 87.74 | Hurt |
| RDIA relabeling of identified data | RDIA, LeNet · MNIST-Label | test accuracy (All, %): 82.46 -> 72.70 | Hurt |
| apply BIR repair to the base model | LeNet · MNIST-Label | test accuracy (All, %): 82.46 -> 95.60 | Improved |
| apply BIR repair to the base model | ResNet50 · CIFAR10-Adv | test accuracy (Noisy classes, %): 79.30 -> 85.85 | Improved |
| apply SGD Influence Repair to the base model | ResNet50 · CIFAR10-Adv | test accuracy (All, %): 87.60 -> 84.08 | Hurt |
| apply BIR repair to the base model | LeNet · MNIST-Input | test accuracy (All, %): 96.46 -> 96.16 | Hurt |
| apply BIR repair to the base model | ResNet20 · CIFAR-10N Worst | test accuracy (%): 77.76 -> 85.50 | Improved |
| apply SGD Influence Repair to the base model | ResNet20 · CIFAR-10N Worst | test accuracy (%): 77.76 -> 71.50 | Hurt |
| update only Layer 1 instead of the selected layers (ablation) | BIR, ResNet20 · CIFAR10-Label | treatment accuracy (%): 83.98 -> 74.82 | Hurt |
| update all layers instead of the selected layer subset (ablation) | BIR, LeNet · MNIST-Label | treatment accuracy (%): 95.60 -> 95.50 | No clear effect |
Closest papers by meaning
- Conserve-Update-Revise to Cure Generalization and Robustness Trade-off in Adversarial Training (ICLR 2024 · similarity 0.85)
- Machine Unlearning Fails to Remove Data Poisoning Attacks (ICLR 2025 · similarity 0.84)
- Outlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Models (ICML 2025 · similarity 0.84)
- The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed Samples (NeurIPS 2025 · similarity 0.84)
- Learning to Unlearn: Instance-Wise Unlearning for Pre-trained Classifiers (AAAI 2024 · similarity 0.83)
- Adversarial Mixup Unlearning (ICLR 2025 · similarity 0.83)
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.
