Research map / Architectures and efficiency / CNN design
Capsule Network Is Not More Robust Than Convolutional Network
Jindong Gu, Volker Tresp, Han Hu
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 |
|---|---|---|---|
| add global average pooling to the ConvNet classifier | ConvNet-Avg · AffNIST | accuracy: 66 -> 94.11 | Improved |
| remove dynamic routing (ablation) | CapsNet · AffNIST | accuracy: 90.07{ ($\pm$ 0.98)} -> 93.55{ ($\pm$ 1.47)} | Improved |
| share the transformation matrix (ablation) | CapsNet without routing · AffNIST | accuracy: 80.49{ ($\pm$ 0.34)} -> 93.55{ ($\pm$ 1.47)} | Improved |
| add the squashing function (ablation) | CapsNet without routing · AffNIST | accuracy: 80.42{ ($\pm$ 0.39)} -> 93.55{ ($\pm$ 1.47)} | Improved |
| use normal rather than conditional reconstruction (ablation) | CapsNet without routing · AffNIST | accuracy: 93.55{ ($\pm$ 1.47)} -> 95.09{ ($\pm$ 0.56)} | Improved |
| replace margin loss with cross-entropy loss (ablation) | CapsNet without routing · AffNIST | accuracy: 93.55{ ($\pm$ 1.47)} -> 94.67{ ($\pm$ 0.43)} | Improved |
| use large kernels in ConvNet-FC | ConvNet-FC(LK) · MultiMNIST | accuracy: 95.18\% -> 97.11\% | Improved |
| remove dynamic routing (ablation) | CapsNet · MultiMNIST | accuracy: 93.26{ ($\pm$ 0.24)} -> 90.07{ ($\pm$ 0.29)} | Hurt |
| use a shared transformation matrix (ablation) | CapsNet · MultiMNIST | accuracy: 93.26{ ($\pm$ 0.24)} -> 86.44{ ($\pm$ 0.37)} | Hurt |
| remove the squashing function (ablation) | CapsNet · MultiMNIST | accuracy: 93.26{ ($\pm$ 0.24)} -> 87.24{ ($\pm$ 0.53)} | Hurt |
| replace conditional reconstruction with normal reconstruction (ablation) | CapsNet · MultiMNIST | accuracy: 93.26{ ($\pm$ 0.24)} -> 93.19{ ($\pm$ 0.30)} | No clear effect |
| remove reconstruction (ablation) | CapsNet · MultiMNIST | accuracy: 93.26{ ($\pm$ 0.24)} -> 90.17{ ($\pm$ 0.26)} | Hurt |
Closest papers by meaning
- Improving the Robustness of Capsule Networks to Image Affine Transformations (CVPR 2020 · similarity 0.91)
- PT-CapsNet: A Novel Prediction-Tuning Capsule Network Suitable for Deeper Architectures (ICCV 2021 · similarity 0.88)
- STAR-Caps: Capsule Networks with Straight-Through Attentive Routing (NeurIPS 2019 · similarity 0.87)
- A Comprehensive Study on Robustness of Image Classification Models: Benchmarking and Rethinking (IJCV 2024 · similarity 0.87)
- OrthCaps: An Orthogonal CapsNet with Sparse Attention Routing and Pruning (CVPR 2024 · similarity 0.86)
- Revealing the Dark Secrets of Extremely Large Kernel ConvNets on Robustness (ICML 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.
