Research map / Adversarial robustness and security / Certified and empirical robustness
Well-Classified Examples Are Underestimated in Classification with Deep Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add an encouraging bonus to CE loss | CNN · MNIST | accuracy: 99.42$\pm$0.06 -> 99.56$\pm$0.05 | Helped |
| Add an encouraging bonus to CE loss | ResNet-50 · CIFAR-10 | accuracy: 92.34$\pm$0.70 -> 92.97$\pm$0.42 | Helped |
| Add an encouraging bonus to CE loss | ResNet-50 · ImageNet | top-1 accuracy: 75.49$\pm$0.28 -> 76.43$\pm$0.15 | Helped |
| Add an encouraging bonus to CE loss | Transformer · IWSLT Fr-En | BLEU: 37.10$\pm$0.06 -> 37.73$\pm$0.17 | Helped |
| Add a bonus to CE loss | ResNet-50 · iNaturalist 2018 | accuracy (ALL): 64.3 -> 65.3 (conservative); 65.8 (normal); 66.3 (aggressive) | Helped |
| Add a bonus to CE loss | Classifier · MNIST vs Fashion-MNIST | FPR95: 20.92 -> 8.69 | Helped |
| Use half the learning rate with EL (ablation) | ResNet-50 with EL · CIFAR-10 | accuracy: 92.74$\pm$0.38 -> 93.69$\pm$0.25 | Helped |
| Use half the learning rate with EL (ablation) | ResNet-50 with EL · CIFAR-100 | accuracy: 75.80$\pm$0.09 -> 76.37$\pm$0.29 | Helped |
| Add mirror bonus to MSE (ablation) | MSE classifier · MNIST | accuracy: 99.56 $\pm$ 0.05 -> 99.66 $\pm$ 0.04 | Helped |
| Add normal bonus to deferred re-weighting (ablation) | Deferred Re-weighting · iNaturalist 2018 | accuracy (ALL): 68.1 -> 70.3 | Helped |
| Add normal bonus to representation learning and classifier re-training (ablation) | Decoupling Reps&Cls (CRT) · iNaturalist 2018 | accuracy (ALL): 64.9 -> 66.8 | Helped |
| Add encouraging loss to large-margin softmax (ablation) | Large-margin softmax · MNIST | accuracy: 99.62$\pm$ 0.04 -> 99.64 $\pm$ 0.02 | Helped |
Closest papers by meaning
- Improving Adversarial Robustness Requires Revisiting Misclassified Examples (ICLR 2020 · similarity 0.86)
- Difficult Examples Hurt Unsupervised Contrastive Learning: A Theoretical Perspective (ICLR 2026 · similarity 0.86)
- Deep Learning on a Data Diet: Finding Important Examples Early in Training (NeurIPS 2021 · similarity 0.86)
- One Man's Trash Is Another Man's Treasure: Resisting Adversarial Examples by Adversarial Examples (CVPR 2020 · similarity 0.86)
- Stable Unlearnable Example: Enhancing the Robustness of Unlearnable Examples via Stable Error-Minimizing Noise (AAAI 2024 · similarity 0.84)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.84)
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.