Research map / Optimization / Implicit bias and generalization
A Deeper Look at the Hessian Eigenspectrum of Deep Neural Networks and its Applications to Regularization
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 HTR to cross-entropy training | LeNet · MNIST | generalization error: 1.25 ± 0.03 -> 1.19 ± 0.07 | Helped |
| Add HTR to cross-entropy training | LeNet · FMNIST | generalization error: 11.22 ± 0.11 -> 11.14 ± 0.04 | Helped |
| Add HTR to cross-entropy training | LeNet · SVHN | generalization error: 10.94 ± 0.06 -> 10.77 ± 0.26 | Helped |
| Add HTR to cross-entropy training | VGG11 · CIFAR10 | generalization error: 18.20 ± 0.45 -> 15.11 ± 0.22 | Helped |
| Add HTR to cross-entropy training | VGG13 · SVHN | generalization error: 7.73 ± 0.28 -> 7.47 ± 0.01 | Helped |
| Add HTR to cross-entropy training | ResNet18 · CIFAR10 | generalization error: 11.97 ± 0.14 -> 11.95 ± 0.22 | No clear effect |
| Add HTR to cross-entropy training | VGG11-BN · CIFAR100 | generalization error: 45.25 ± 0.44 -> 44.98 ± 0.05 | Helped |
| Apply HTR only to middle layers rather than all layers (ablation) | LeNet · MNIST | generalization error: 1.19 ± 0.07 -> 1.18 ± 0.07 | Helped |
| Apply HTR only to middle layers rather than all layers (ablation) | LeNet · FMNIST | generalization error: 11.14 ± 0.04 -> 11.10 ± 0.07 | Helped |
| Apply HTR only to middle layers rather than all layers (ablation) | ResNet18 · CIFAR10 | generalization error: 11.95 ± 0.22 -> 11.87 ± 0.12 | Helped |
| Apply HTR only to middle layers rather than all layers (ablation) | VGG11-BN · CIFAR100 | generalization error: 44.98 ± 0.05 -> 44.81 ± 0.15 | Helped |
| Combine HTR with L2 regularization (ablation) | HTR-trained models · MNIST, FMNIST, SVHN, CIFAR10, and CIFAR100 | generalization error | Helped |
Closest papers by meaning
- Investigating the Overlooked Hessian Structure: From CNNs to LLMs (ICML 2025 · similarity 0.91)
- Loss Landscape Characterization of Neural Networks without Over-Parametrization (NeurIPS 2024 · similarity 0.87)
- Super Consistency of Neural Network Landscapes and Learning Rate Transfer (NeurIPS 2024 · similarity 0.87)
- Neglected Hessian component explains mysteries in sharpness regularization (NeurIPS 2024 · similarity 0.87)
- Exploring The Loss Landscape Of Regularized Neural Networks Via Convex Duality (ICLR 2025 · similarity 0.87)
- Simplicity Bias via Global Convergence of Sharpness Minimization (ICML 2024 · similarity 0.87)
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.