Research map / Adversarial robustness and security / Certified and empirical robustness
Group-Wise Dynamic Dropout Based on Latent Semantic Variations
Zhiwei Ke, Zhiwei Wen, Weicheng Xie, Yi Wang, Linlin Shen
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 |
|---|---|---|---|
| Replace original dropout with FDD-1D on FC features | ResNet-18 dropout · CIFAR-100 | accuracy (%): 75.59 -> 76.81 | Improved |
| Replace original dropout with FDD-1D | ResNet-18 dropout · Fashion-MNIST | accuracy (%): 93.97 -> 94.33 | Improved |
| Replace original dropout with FDD-1D | ResNet-18 dropout · CIFAR-10 | accuracy (%): 94.24 -> 94.65 | Improved |
| Replace original dropout with FDD-1D | ResNet-18 dropout · SVHN | accuracy (%): 96.52 -> 96.85 | Improved |
| Replace spatial dropout with FDD-2D on feature maps | ResNet-18 dropout · CIFAR-100 | accuracy (%): 75.13 -> 76.83 | Improved |
| Use FDD-3D instead of FDD-2D (3D PCA projection) (ablation) | ResNet-18 dropout · CIFAR-100 | accuracy (%): 76.83 -> 76.53 | No clear effect |
| Correlate dropout probability positively instead of negatively with feature density (ablation) | ResNet-18 dropout · Fashion-MNIST, CIFAR-10, CIFAR-100, SVHN | accuracy (%) | Hurt |
| Correlate dropout probability negatively with feature density (ablation) | ResNet-18 dropout · Fashion-MNIST, CIFAR-10, CIFAR-100, SVHN | accuracy (%) | Improved |
| Add FDD-1D dropout under FGSM attack | ResNet-18 dropout · CIFAR-100 | adversarial accuracy (%): 15 -> 66.28 | Improved |
| Add FDD-1D dropout under FGSM attack | ResNet-18 dropout · Fashion-MNIST | adversarial accuracy (%): 5.85 -> 39.62 | Improved |
| Use original random dropout instead of no dropout under attack | ResNet-18 dropout · CIFAR-100 | adversarial accuracy (%): 14.72 -> 15 | No clear effect |
| Use spatial dropout instead of no dropout under attack | ResNet-18 dropout · Fashion-MNIST | adversarial accuracy (%): 11.95 -> 9.93 | No clear effect |
Closest papers by meaning
- Advanced Dropout: A Model-free Methodology for Bayesian Dropout Optimization (TPAMI 2021 · similarity 0.89)
- Surprising properties of dropout in deep networks (JMLR 2018 · similarity 0.89)
- Demystifying Dropout (ICML 2019 · similarity 0.88)
- Curriculum Dropout (ICCV 2017 · similarity 0.88)
- Adversarial Dropout for Supervised and Semi-Supervised Learning (AAAI 2018 · similarity 0.88)
- Dropout Reduces Underfitting (ICML 2023 · similarity 0.88)
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.
