Research map / Architectures and efficiency / CNN design
Capsules with Inverted Dot-Product Attention Routing
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 Inverted Dot-Product Attention Routing instead of Dynamic Routing | CapsNet with simple backbone · CIFAR-10 | test accuracy: 84.08% -> 85.17 | Helped |
| use Inverted Dot-Product Attention Routing instead of EM Routing | CapsNet with simple backbone · CIFAR-100 | test accuracy: 37.73% -> 57.32% | Helped |
| use Inverted Dot-Product Attention Routing instead of Dynamic Routing | CapsNet with ResNet backbone · CIFAR-10 | test accuracy: 92.65% -> 95.14% | Helped |
| use Inverted Dot-Product Attention Routing instead of Dynamic Routing | CapsNet with ResNet backbone · CIFAR-100 | test accuracy: 71.70% -> 78.02% | Helped |
| use Inverted Dot-Product Attention Routing with ResNet instead of ResNet-18 | CIFAR classification models · CIFAR-10 | test accuracy: 95.11% -> 95.14% | No clear effect |
| remove Layer Normalization (ablation) | Inverted Dot-Product Attention Routing · CIFAR-10 | test accuracy | Hurt |
| add Layer Normalization (ablation) | Inverted Dot-Product Attention Routing · CIFAR-10 | training stability | Helped |
| replace concurrent routing with sequential iterative routing (ablation) | Inverted Dot-Product Attention Routing · CIFAR-10 | test accuracy across routing iterations | Hurt |
| add capsule activations (ablation) | Inverted Dot-Product Attention Routing · CIFAR-10 | test accuracy | Hurt |
| use Inverted Dot-Product Attention Routing instead of Dynamic Routing | CapsNet with vector poses · DiverseMultiMNIST | test accuracy: 83.39% to 85.74% | Helped |
| replace CNN average pooling with capsule routing | same-depth and same-width image classifier · DiverseMultiMNIST | test accuracy: 79.81% -> 85.74% | Helped |
| replace average pooling with a learned fully connected operation | Baseline CNN · DiverseMultiMNIST | test accuracy: 79.81% -> 84.84% | Helped |
Closest papers by meaning
- A Receptor Skeleton for Capsule Neural Networks (ICML 2021 · similarity 0.89)
- Self-Routing Capsule Networks (NeurIPS 2019 · similarity 0.89)
- Dynamic Routing Between Capsules (NeurIPS 2017 · similarity 0.89)
- STAR-Caps: Capsule Networks with Straight-Through Attentive Routing (NeurIPS 2019 · similarity 0.88)
- OrthCaps: An Orthogonal CapsNet with Sparse Attention Routing and Pruning (CVPR 2024 · similarity 0.86)
- ParseCaps: An Interpretable Parsing Capsule Network for Medical Image Diagnosis (AAAI 2025 · 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.