Research map / Architectures and efficiency / CNN design
SKFAC: Training Neural Networks With Faster Kronecker-Factored Approximate Curvature
Check 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 |
|---|---|---|---|
| Replace K-FAC with low-rank K-FAC using Reduce Sum | ResNet-50 · ImageNet-1K | training time to reach 75% validation accuracy: 1544.9 -> 1189.5 min; -23.00% | Helped |
| Replace K-FAC with low-rank K-FAC using Spatial Subsampling | ResNet-50 · ImageNet-1K | training time to reach 75% validation accuracy: 1544.9 -> 1206.9 min; -21.88% | Helped |
| Replace K-FAC with low-rank K-FAC using Reduce Sum | ResNet-50 · ImageNet-1K | epochs to reach 75% validation accuracy: 33 -> 36 | Hurt |
| Replace K-FAC with low-rank K-FAC using Spatial Subsampling | ResNet-50 · ImageNet-1K | epochs to reach 75% validation accuracy: 33 -> 37 | Hurt |
| Use SUA for convolutional-layer activations | K-FAC on ResNet-50 · ImageNet-1K | training time to reach 75% validation accuracy: 1544.9 -> 1494.5 min; -3.3% | Helped |
| Vary Spatial Subsampling amount across 1, 2, 4, and 8 (ablation) | SKFAC-SS · CIFAR-10 | test accuracy | No clear effect |
Closest papers by meaning
- A Kronecker-factored approximate Fisher matrix for convolution layers (ICML 2016 · similarity 0.93)
- Kronecker-Factored Approximate Curvature for Modern Neural Network Architectures (NeurIPS 2023 · similarity 0.93)
- Kronecker-Factored Approximate Curvature for Physics-Informed Neural Networks (NeurIPS 2024 · similarity 0.92)
- Structured Inverse-Free Natural Gradient Descent: Memory-Efficient & Numerically-Stable KFAC (ICML 2024 · similarity 0.86)
- Convolutions and More as Einsum: A Tensor Network Perspective with Advances for Second-Order Methods (NeurIPS 2024 · similarity 0.85)
- Global curvature for second-order optimization of neural networks (ICML 2025 · 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.