Research map / Learning theory and algorithms / Matrix and tensor methods
EA-CG: An Approximate Second-Order Method for Training Fully-Connected Neural Networks
Sheng-Wei Chen, Chun-Nan Chou, Edward Y. Chang
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 |
|---|---|---|---|
| Use PCH curvature with EA-CG instead of KFI | FCNN training · Cifar-10, cross-entropy | testing accuracy | Improved |
| Use EA-CG instead of KFI | FCNN training · Cifar-10, cross-entropy | wall-clock convergence | Improved |
| Use EA-CG instead of KFI | FCNN training · ImageNet-10, cross-entropy | wall-clock time | No clear effect |
| Use PCH-1 or PCH-2 rather than Fisher or Gauss-Newton curvature | FCNN Hessian approximation · Cifar-10, cross-entropy | total Hessian error: 0.1535 vs 0.0402; 0.0446 vs 0.0330 | Improved |
| Use PCH-1 or PCH-2 rather than Fisher or Gauss-Newton curvature | FCNN Hessian approximation · ImageNet-10, cross-entropy | total Hessian error: 0.0910 vs 0.0424; 0.0436 vs 0.0369 | Improved |
| Use PCH-1 rather than Fisher curvature | FCNN Hessian approximation · Cifar-10, non-convex criterion | total Hessian error: 0.1198 vs 0.0349 | Improved |
| Use PCH-1 rather than Fisher curvature | FCNN Hessian approximation · ImageNet-10, non-convex criterion | total Hessian error: 0.0880 vs 0.0598 | Improved |
| Use the PCH-1 eigenvalue treatment rather than PCH-2 (ablation) | FCNN Hessian approximation · Cifar-10, cross-entropy | total Hessian error: 0.0402 vs 0.0330 | Hurt |
| Use the PCH-1 eigenvalue treatment rather than PCH-2 (ablation) | FCNN Hessian approximation · ImageNet-10, cross-entropy | total Hessian error: 0.0424 vs 0.0369 | Hurt |
| Use the PCH-1 eigenvalue treatment rather than PCH-2 (ablation) | FCNN Hessian approximation · Cifar-10, non-convex criterion | total Hessian error: 0.0853 vs 0.0349 | Improved |
| Use the PCH-1 eigenvalue treatment rather than PCH-2 (ablation) | FCNN Hessian approximation · ImageNet-10, non-convex criterion | total Hessian error: 0.0660 vs 0.0598 | Improved |
| Use PCH-1 instead of PCH-2 (ablation) | FCNN training · Cifar-10 and ImageNet-10, convex and non-convex criteria | training loss and testing accuracy | No clear effect |
Closest papers by meaning
- Practical Quasi-Newton Methods for Training Deep Neural Networks (NeurIPS 2020 · similarity 0.87)
- A Layer-Wise Natural Gradient Optimizer for Training Deep Neural Networks (NeurIPS 2024 · similarity 0.84)
- Learning Sparse Approximate Inverse Preconditioners for Conjugate Gradient Solvers on GPUs (NeurIPS 2025 · similarity 0.84)
- Convergence Analysis of Newton's Method for Neural Networks in the Overparameterized Limit (NeurIPS 2026 · similarity 0.83)
- PDHG-Unrolled Learning-to-Optimize Method for Large-Scale Linear Programming (ICML 2024 · similarity 0.82)
- Provable Acceleration of Nesterov's Accelerated Gradient for Asymmetric Matrix Factorization and Linear Neural Networks (NeurIPS 2024 · similarity 0.82)
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.
