Research map / Adversarial robustness and security / Certified and empirical robustness
Generalist: Decoupling Natural and Robust Generalization
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 |
|---|---|---|---|
| Decouple natural and adversarial training into task-specific learners and periodically combine their parameters | ResNet-18 on CIFAR-10 · CIFAR-10 | natural accuracy: 84.32 -> 89.09 | Helped |
| Decouple natural and adversarial training into task-specific learners and periodically combine their parameters | WRN-32-10 on CIFAR-10 · CIFAR-10 | AA accuracy: 52.19 -> 52.91 | Helped |
| Decouple natural and adversarial training into task-specific learners and periodically combine their parameters | ResNet-18 on CIFAR-100 · CIFAR-100 | AA accuracy: 23.72 -> 23.96 | Helped |
| Add weight averaging to both natural and adversarial base learners (ablation) | Generalist on CIFAR-10 · CIFAR-10 | accuracy on corresponding tasks | Helped |
| Add weight averaging only to the natural base learner (ablation) | Generalist on CIFAR-10 · CIFAR-10 | robust accuracy | Hurt |
| Use Adam for the adversarial base learner instead of SGD (ablation) | Generalist on CIFAR-10 · CIFAR-10 | robust accuracy | Hurt |
| Schedule the mixing ratio instead of using a static γ=0.5 (ablation) | Generalist on CIFAR-10 · CIFAR-10 | robust accuracy | Helped |
| Use γ=(1,1,1,0) with communication frequency c=5 after epoch 75 (ablation) | Generalist on CIFAR-10 · CIFAR-10 | natural and robust accuracy | Helped |
| Vary communication frequency c (ablation) | Generalist on CIFAR-10 · CIFAR-10 | natural and robust accuracy | Mixed |
| Set adversarial learning rate to 0.1 rather than 0.01 while keeping natural learning rate at 0.1 (ablation) | Generalist on CIFAR-10 · CIFAR-10 | AA accuracy: 46.37 -> 41.86 | Hurt |
| Set adversarial learning rate to 0.001 rather than 0.01 while keeping natural learning rate at 0.1 (ablation) | Generalist on CIFAR-10 · CIFAR-10 | natural accuracy: 89.09 -> 90.45 | Helped |
Closest papers by meaning
- Certified Robust Neural Networks: Generalization and Corruption Resistance (ICML 2023 · similarity 0.87)
- Stability and Generalization of Adversarial Training for Shallow Neural Networks with Smooth Activation (NeurIPS 2024 · similarity 0.87)
- Attacks Which Do Not Kill Training Make Adversarial Learning Stronger (ICML 2020 · similarity 0.87)
- Improved OOD Generalization via Adversarial Training and Pretraing (ICML 2021 · similarity 0.86)
- To Grok or not to Grok: Disentangling Generalization and Memorization on Corrupted Algorithmic Datasets (ICLR 2024 · similarity 0.85)
- Exact, Tractable Gauss-Newton Optimization in Deep Reversible Architectures Reveal Poor Generalization (NeurIPS 2024 · similarity 0.85)
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.