Research map / Architectures and efficiency / RNNs and normalization
Fast Certified Robust Training with Short Warmup
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 BN after the fully-connected layer (BN-FC) (ablation) | Vanilla IBP · CIFAR-10 eps=8/255, CNN-7, 160 epochs | verified error (%): 69.21 ± 0.30 -> 67.01 ± 0.29 | Helped |
| add BN after convolutions only (ablation) | Vanilla IBP · CIFAR-10 eps=8/255, CNN-7, 160 epochs | verified error (%): 69.43 ± 0.28 -> 69.21 ± 0.30 | No clear effect |
| add IBP initialization without regularizers (ablation) | IBP with full BN · CIFAR-10 eps=8/255, CNN-7, 160 epochs | verified error (%): 67.01 ± 0.29 -> 67.63 ± 0.34 | Hurt |
| add tightness regularizer on top of init (ablation) | IBP with full BN + IBP init · CIFAR-10 eps=8/255, CNN-7, 160 epochs | verified error (%): 67.63 ± 0.34 -> 66.74 ± 0.20 | Helped |
| add ReLU balance regularizer on top of init (ablation) | IBP with full BN + IBP init · CIFAR-10 eps=8/255, CNN-7, 160 epochs | verified error (%): 67.63 ± 0.34 -> 66.46 ± 0.44 | Helped |
| add IBP initialization to full method (ablation) | Full method w/ BN and both regularizers · CIFAR-10 eps=8/255, CNN-7, 160 epochs | verified error (%): 66.34 ± 0.38 -> 65.58 ± 0.32 | Helped |
| add IBP initialization to full method (short schedule) (ablation) | Full method · CIFAR-10 eps=8/255, CNN-7, 70 epochs | verified error (%): 69.29 ± 0.28 -> 68.81 ± 0.24 | Helped |
| full proposed method vs Vanilla IBP (ablation) | Vanilla IBP · CIFAR-10 eps=8/255, CNN-7, 160 epochs | verified error (%): 67.01 ± 0.29 -> 65.58 ± 0.32 | Helped |
| full proposed method vs CROWN-IBP with loss fusion | CROWN-IBP · CIFAR-10 eps=8/255, CNN-7, 160 epochs | verified error (%): 69.67 ± 0.38 -> 65.58 ± 0.32 | Helped |
| full proposed method vs Vanilla IBP | Vanilla IBP · TinyImageNet eps=1/255, CNN-7, 80 epochs (1+20+59) | verified error (%): 82.84 -> 82.36 | Helped |
| use Kaiming instead of Xavier initialization (ablation) | Vanilla IBP · CIFAR-10 eps=8/255, CNN-7, 160 epochs | verified error (%): 67.01 -> 68.07 | Hurt |
| remove centralization and unitization from BN (ablation) | Full method · CIFAR-10 eps=8/255, CNN-7, 160 epochs | verified error (%): 65.58 ± 0.30 -> 69.91 ± 0.31 | Hurt |
Closest papers by meaning
- Understanding Certified Training with Interval Bound Propagation (ICLR 2024 · similarity 0.90)
- Towards Stable and Efficient Training of Verifiably Robust Neural Networks (ICLR 2020 · similarity 0.88)
- Analyzing & Reducing the Need for Learning Rate Warmup in GPT Training (NeurIPS 2024 · similarity 0.82)
- CTBench: A Library and Benchmark for Certified Training (ICML 2025 · similarity 0.82)
- Scaling FP8 training to trillion-token LLMs (ICLR 2025 · similarity 0.81)
- SDP-CROWN: Efficient Bound Propagation for Neural Network Verification with Tightness of Semidefinite Programming (ICML 2025 · similarity 0.81)
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.