Research map / Architectures and efficiency / RNNs and normalization
Towards Faster Training of Global Covariance Pooling Networks by Iterative Matrix Square Root Normalization
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 |
|---|---|---|---|
| use iterative Newton-Schulz square-root normalization instead of EIG-based MPN-COV | ImageNet AlexNet covariance pooling · ImageNet val | top-1 error (%): 38.51 -> 38.45 | Helped |
| use iterative Newton-Schulz square-root normalization instead of EIG-based MPN-COV | ImageNet ResNet-50 covariance pooling · ImageNet val | top-1 error (%): 22.73 -> 22.14 | Helped |
| use trace rather than Frobenius norm for pre-normalization (ablation) | iSQRT-COV with AlexNet · ImageNet val | top-1 error (%): 38.78 -> 38.45 | Helped |
| replace trace post-compensation with batch normalization (ablation) | iSQRT-COV with ResNet-50 · ImageNet val | top-1 error (%): 22.14 -> 23.12 | Hurt |
| remove post-compensation (ablation) | iSQRT-COV with ResNet-50 · ImageNet val | top-1 error (%) | Hurt |
| increase Newton-Schulz iteration count beyond 7 (ablation) | iSQRT-COV with AlexNet · ImageNet val | top-1 error rate | Hurt |
| reduce the representation dimension from 32K to 8K (ablation) | iSQRT-COV with ResNet-50 · ImageNet val | top-1 error (%): 22.14 -> 22.78 | Hurt |
| reduce the representation dimension from 32K to 2K (ablation) | iSQRT-COV with ResNet-50 · ImageNet val | top-1 error (%): 22.14 -> 23.73 | Hurt |
| reduce representation dimension from 32K to 8K (ablation) | iSQRT-COV with ResNet-50 · CUB-200-2011 | accuracy (%): 88.1 -> 87.3 | Hurt |
| reduce representation dimension from 32K to 8K (ablation) | iSQRT-COV with ResNet-50 · FGVC-aircraft | accuracy (%): 90.0 -> 89.5 | Hurt |
| reduce representation dimension from 32K to 8K (ablation) | iSQRT-COV with ResNet-50 · Stanford Cars | accuracy (%): 92.8 -> 91.7 | Hurt |
| increase training from 60 to 90 epochs (ablation) | iSQRT-COV with ResNet-50 · ImageNet val | top-5 error (%): 6.22 -> 6.12 | Helped |
Closest papers by meaning
- Revitalizing SVD for Global Covariance Pooling: Halley’s Method to Overcome Over-Flattening (NeurIPS 2025 · similarity 0.90)
- DropCov: A Simple yet Effective Method for Improving Deep Architectures (NeurIPS 2022 · similarity 0.89)
- Global Second-Order Pooling Convolutional Networks (CVPR 2019 · similarity 0.87)
- What Deep CNNs Benefit From Global Covariance Pooling: An Optimization Perspective (CVPR 2020 · similarity 0.87)
- Understanding Matrix Function Normalizations in Covariance Pooling through the Lens of Riemannian Geometry (ICLR 2025 · similarity 0.85)
- Geometry-Aware Edge Pooling for Graph Neural Networks (NeurIPS 2025 · similarity 0.84)
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.