Research map / Architectures and efficiency / CNN design
Scaling-up Diverse Orthogonal Convolutional Networks by a Paraunitary Framework
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 SC-Fac instead of previous orthogonal-convolution designs | Standard convolution layers · Standard convolution | orthogonality precision: (+3.14 \pm 7.38) \times 10^{-8} vs (+2.88 \pm 1.90) \times 10^{-4} (CayleyConv) | Helped |
| Use SC-Fac for dilated, strided, and grouped convolution variants | SC-Fac · Dilated, strided, and grouped convolutions | orthogonality precision | Helped |
| Use an additive skip connection rather than concatenation or no skip connection (ablation) | Deep Lipschitz networks · Deep Lipschitz networks | accuracy | Helped |
| Remove skip connections (ablation) | Deep Lipschitz networks · Deep Lipschitz networks | accuracy | Hurt |
| Increase network depth or width (ablation) | Orthogonal Lipschitz networks · Deep and wide architectures | clean and robust accuracy | Helped |
| Use identical initialization rather than other tested initializations (ablation) | Deep Lipschitz networks · Networks with more than 10 layers | accuracy | Helped |
| Change initialization method among identical, permutation, uniform, and torus (ablation) | Shallow Lipschitz networks · Shallow networks | accuracy | No clear effect |
| Use average pooling instead of strided convolution for down-sampling (ablation) | Orthogonal Lipschitz networks · Down-sampling comparison | accuracy | Helped |
| Increase kernel size (ablation) | Orthogonal Lipschitz networks · Kernel and dilation study | accuracy | Helped |
| Use SC-Fac rather than CayleyConv or BCOP | Shallow Lipschitz networks · CIFAR-10 certified and practical robustness | robust accuracy | No clear effect |
| Use SC-Fac rather than other convolution designs | WideResNet22 Lipschitz network · CIFAR-10 practical robustness | clean and robust accuracy | Helped |
Closest papers by meaning
- Orthogonal Convolutional Neural Networks (CVPR 2020 · similarity 0.92)
- An Adaptive Orthogonal Convolution Scheme for Efficient and Flexible CNN Architectures (ICML 2025 · similarity 0.91)
- Constructing Orthogonal Convolutions in an Explicit Manner (ICLR 2022 · similarity 0.90)
- Deeply Shared Filter Bases for Parameter-Efficient Convolutional Neural Networks (NeurIPS 2021 · similarity 0.88)
- Revisiting Residual Connections: Orthogonal Updates for Stable and Efficient Deep Networks (NeurIPS 2025 · similarity 0.87)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · 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.