Research map / Label-efficient and robust learning / Recognition and long-tail
Decoupled 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 inner-product convolution with SegConv using cosine angular activation | ResNet-32 · CIFAR-100 | testing error rate (%): 26.69 -> 20.91 | Helped |
| replace standard convolution with LinearConv and square-cosine activation | ResNet-32 · CIFAR-100 | testing error rate (%): 25.16 -> 20.23 | Helped |
| use TanhConv instead of baseline convolution | modified ResNet-18 · ImageNet-2012 with BN | center-crop Top-5 error (%): 12.10 -> 11.10 | Helped |
| use TanhConv instead of baseline convolution | CNN-9 · CIFAR-10, white-box natural training, no attack | accuracy (%): 85.35 -> 91.45 | Helped |
| use decoupled operators instead of the baseline | CNN-9 · CIFAR-10, white-box natural training, FGSM | accuracy (%): 18.82 -> 52.60 | Helped |
| use SphereConv instead of the baseline | CNN-9 · CIFAR-10, black-box natural training, FGSM | accuracy (%): 50.90 -> 56.71 | Helped |
| include weight norm in the TanhConv magnitude function (ablation) | TanhConv · CIFAR-100 | error: 21.17 -> 22.95; 23.03; 23.38 | Hurt |
| replace standard gradients with weight projection (ablation) | TanhConv · CIFAR-100 | error: 23.09 -> 21.17 | Helped |
| replace standard gradients with weighted gradients (ablation) | TanhConv · CIFAR-100 | error: 23.09 -> 21.45 | Helped |
| use square-cosine angular activation without ReLU (ablation) | SegConv in CNN-9 · CIFAR-100 | testing error rate (%): 26.01 -> 24.29 | Helped |
| train with SGD instead of ADAM (ablation) | ResNet baseline · CIFAR-100 | testing error rate (%): 26.69 -> 21.55 | Helped |
| use SphereConv instead of baseline convolution | CNN-9 · CIFAR-10, black-box attacks with natural training | accuracy (%): 50.90 -> 56.71; 36.22 -> 43.10 | Helped |
Closest papers by meaning
- Decoupled Convolutions for CNNs (AAAI 2018 · similarity 0.89)
- Network Deconvolution (ICLR 2020 · similarity 0.86)
- Inverting Visual Representations With Convolutional Networks (CVPR 2016 · similarity 0.86)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.84)
- Large Vision-Language Models Get Lost in Attention (ICML 2026 · similarity 0.84)
- Catch-Up Mix: Catch-Up Class for Struggling Filters in CNN (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.