Research map / Architectures and efficiency / RNNs and normalization
Orthogonal Weight Normalization: Solution to Optimization Over Multiple Dependent Stiefel Manifolds in Deep Neural Networks
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 |
|---|---|---|---|
| Replace convolution layers with OLM | VGG-style · CIFAR-10 | test error (%): 10.39 ± 0.14 -> 8.61 ± 0.18 | Helped |
| Replace convolution layers with OLM | VGG-style · CIFAR-100 | test error (%): 36.02 ± 0.40 -> 32.58 ± 0.10 | Helped |
| Replace convolution layers with OLM | BN-Inception · CIFAR-10 | test error (%): 5.38 ± 0.18 -> 4.74 ± 0.16 | Helped |
| Replace convolution layers with OLM | BN-Inception · CIFAR-100 | test error (%): 24.87 ± 0.15 -> 22.02 ± 0.13 | Helped |
| Replace all convolution layers with OLM | WRN-28-10 · CIFAR-10 | test error (%): 4.17 -> 3.73 ± 0.12 | Helped |
| Replace all convolution layers with OLM | WRN-28-10 · CIFAR-100 | test error (%): 20.04 -> 18.76 ± 0.40 | Helped |
| Replace only the first convolution layer with OLM (ablation) | WRN-28-10 · CIFAR-10 | test error (%): 4.17 -> 3.82 ± 0.19 | Helped |
| Replace only the first convolution layer with OLM (ablation) | WRN-28-10 · CIFAR-100 | test error (%): 20.04 -> 18.61 ± 0.14 | Helped |
| Replace first 2 convolution layers with OLM instead of all layers (ablation) | VGG-style · CIFAR-10 | test error (%): 8.61 ± 0.18 -> 10.06 ± 0.23 | Hurt |
| Replace first 2 convolution layers with OLM instead of all layers (ablation) | VGG-style · CIFAR-100 | test error (%): 32.58 ± 0.10 -> 35.42 ± 0.32 | Hurt |
| Replace first 4 convolution layers with OLM instead of all layers (ablation) | VGG-style · CIFAR-10 | test error (%): 8.61 ± 0.18 -> 9.61 ± 0.23 | Hurt |
| Replace first 4 convolution layers with OLM instead of all layers (ablation) | VGG-style · CIFAR-100 | test error (%): 32.58 ± 0.10 -> 33.66 ± 0.11 | Hurt |
Closest papers by meaning
- On orthogonality and learning recurrent networks with long term dependencies (ICML 2017 · similarity 0.88)
- Controllable Orthogonalization in Training DNNs (CVPR 2020 · similarity 0.87)
- Orthogonal Over-Parameterized Training (CVPR 2021 · similarity 0.86)
- Feedback Gradient Descent: Efficient and Stable Optimization with Orthogonality for DNNs (AAAI 2022 · similarity 0.86)
- HADAMRNN: BINARY AND SPARSE TERNARY ORTHOGONAL RNNS (ICLR 2025 · similarity 0.84)
- Simple Orthogonal Graph Representation Learning (Student Abstract) (AAAI 2024 · 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.