Research map / Architectures and efficiency / CNN design
Improving the Robustness of Capsule Networks to Image Affine Transformations
Jindong Gu, Volker Tresp
Check your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| remove routing and use uniform coupling coefficients (ablation) | CapsNet · MNIST, FMNIST, SVHN, and CIFAR10 test sets | accuracy: MNIST: 99.41{ ($\pm$ 0.08)} -> 99.54{ ($\pm$ 0.04)}; FMNIST: 92.12{ ($\pm$ 0.29)} -> 92.53{ ($\pm$ 0.26)}; SVHN: 91.32{ | No clear effect |
| detach coupling coefficients during backpropagation rather than unroll routing iterations (ablation) | CapsNet · MNIST, FMNIST, SVHN, and CIFAR10 test sets | accuracy: MNIST: 99.41{ ($\pm$ 0.08)} -> 99.29{ ($\pm$ 0.09)}; FMNIST: 92.12{ ($\pm$ 0.29)} -> 91.53{ ($\pm$ 0.22)}; SVHN: 91.32{ | No clear effect |
| replace dynamic routing with one-step routing using trainable coupling coefficients | CapsNet · MNIST, FMNIST, SVHN, and CIFAR10 test sets | accuracy: MNIST: 99.41{ ($\pm$ 0.08)} -> 99.55{ ($\pm$ 0.04)}; FMNIST: 92.12{ ($\pm$ 0.29)} -> 92.58{ ($\pm$ 0.10)}; SVHN: 91.32{ | No clear effect |
| remove routing and use uniform coupling coefficients (ablation) | CapsNet · AffNIST test | accuracy: 79\% -> 81.81\% | Improved |
| share the transformation matrix across primary capsules | Aff-CapsNet + No-R · AffNIST test | accuracy: 79\% -> 93.21{ ($\pm$0.65)}\% | Improved |
| remove routing from a CapsNet and keep its parallel capsule attention behavior (ablation) | CapsNet · MultiMNIST test | accuracy: 95\% -> 95,49\% | No clear effect |
Closest papers by meaning
- Capsule Network Is Not More Robust Than Convolutional Network (CVPR 2021 · similarity 0.91)
- PT-CapsNet: A Novel Prediction-Tuning Capsule Network Suitable for Deeper Architectures (ICCV 2021 · similarity 0.89)
- Self-Routing Capsule Networks (NeurIPS 2019 · similarity 0.88)
- Matrix capsules with EM routing (ICLR 2018 · similarity 0.87)
- Ladder Capsule Network (ICML 2019 · similarity 0.87)
- OrthCaps: An Orthogonal CapsNet with Sparse Attention Routing and Pruning (CVPR 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.
