Research map / Adversarial robustness and security / Certified and empirical robustness
Hessian-based Analysis of Large Batch Training and Robustness to Adversaries
Zhewei Yao, Amir Gholami, Qi Lei, Kurt Keutzer, Michael W. Mahoney
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 |
|---|---|---|---|
| Increase C1 training batch size from 16 to 2048 | C1 · CIFAR-10 test | accuracy: 77.68 -> 77.31 | Mixed |
| Increase C1 training batch size from 16 to 2048 | C1 · CIFAR-10 test, ε=0.02 | adversarial accuracy: 30.38 -> 3.77 | Hurt |
| Increase CR training batch size from 64 to 256 | CR · CIFAR-100 test | accuracy: 70.81 -> 68.6 | Hurt |
| Increase C1 training batch size from 256 to 16384 | C1 · SVHN test | accuracy: 95.70 -> 94.86 | Hurt |
| Use FGSM adversarial training instead of ordinary training | MNIST model · MNIST adversarial samples, mean of attacks | accuracy: 53.44 -> 87.59 | Improved |
| Use FGSM adversarial training instead of ordinary training | CIFAR-10 model · CIFAR-10 adversarial samples, mean of attacks | accuracy: 16.93 -> 62.51 | Improved |
| Use FGSM adversarial training instead of ordinary training | MNIST model · MNIST clean test | accuracy: 99.32 -> 99.49 | No clear effect |
| Replace ReLU with Softplus (β=20) (ablation) | C1 · CIFAR-10 test, batch 128 vs 2048 | accuracy: 78.79 -> 76.27 | Hurt |
| Replace ReLU with ELU (α=1) (ablation) | C1 · CIFAR-10 test, batch 128 vs 2048 | accuracy: 78.94 -> 76.64 | Hurt |
| Replace ReLU with Softplus (β=20) (ablation) | C1 · CIFAR-10 adversarial test, ε=0.02, batch 128 vs 2048 | adversarial accuracy: 17.79 -> 2.29 | Hurt |
| Replace ReLU with ELU (α=1) (ablation) | C1 · CIFAR-10 adversarial test, ε=0.02, batch 128 vs 2048 | adversarial accuracy: 15.26 -> 3.99 | Hurt |
Closest papers by meaning
- Learning Rates as a Function of Batch Size: A Random Matrix Theory Approach to Neural Network Training (JMLR 2022 · similarity 0.90)
- Robust Fine-Tuning of Deep Neural Networks with Hessian-based Generalization Guarantees (ICML 2022 · similarity 0.88)
- Flatland: The Adventures of Gradient Descent with Large Step Sizes (ICML 2026 · similarity 0.88)
- Analytic Insights into Structure and Rank of Neural Network Hessian Maps (NeurIPS 2021 · similarity 0.87)
- Investigating the Overlooked Hessian Structure: From CNNs to LLMs (ICML 2025 · similarity 0.87)
- Enhancing Noise-Robust Losses for Large-Scale Noisy Data Learning (AAAI 2025 · similarity 0.86)
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.
