Research map / Label-efficient and robust learning / Recognition and long-tail
A Receptor Skeleton for Capsule Neural 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 iterative routing with receptor/transmitter matching plus receptor skeleton | CapsNet with simple backbone · CIFAR-10 | accuracy: +2.4 over IDPA-Routing (85.0% to 87.4%) | Helped |
| Replace iterative routing with receptor/transmitter matching plus receptor skeleton | CapsNet with simple backbone · AffNIST (trained on MNIST only) | accuracy: +2.9 over EM Routing... (96.7% vs 93.8%) | Helped |
| Replace iterative routing with receptor/transmitter matching plus receptor skeleton | CapsNet with simple backbone · SmallNorb | accuracy: 97.2% vs EM Routing 98.0% (slightly lower) | No clear effect |
| Replace iterative routing with receptor/transmitter matching plus receptor skeleton | CapsNet with ResNet-18 backbone · CIFAR-100 | accuracy: +0.5 over IDPA-Routing (72.3% to 72.8%) | Helped |
| Replace iterative routing with receptor/transmitter matching plus receptor skeleton | CapsNet with ResNet-18 backbone · CIFAR-10 | accuracy: 93.7% vs ResNet-18 backbone alone 95.1% | No clear effect |
| Replace Dual/Dynamic routing with receptor skeleton in U-shape segmentation model | Matrix capsule U-shape model · JSRT overlapped object segmentation | mean Dice (%): 93.72 vs 92.68 (M-Dual) | Helped |
| Replace routing with receptor skeleton on the vector capsule version | Vector capsule U-shape model · JSRT overlapped object segmentation | mean Dice (%): 92.92 vs 91.88 (V-Dual) and 89.80 (V-Dynamic) | Helped |
| Replace iterative routing with receptor skeleton | CapsNet with simple backbone · computing complexity (inference time and GPU memory) | inference time (fps) and GPU memory: 5.5x to 1600x faster inference; 22% to 83% of GPU memory | Helped |
| Use 2 receptors per capsule instead of 1 (ablation) | CapsNet with simple backbone · CIFAR-10 (K=6) | accuracy: +0.6 (86.8% to 87.4%) | Helped |
| Increase receptors per capsule from 2 to 3 or 6 (ablation) | CapsNet with simple backbone · CIFAR-10 | accuracy: M=3 87.5% vs M=2 87.4%; M>2 fluctuates slightly | No clear effect |
| Vary K nearest neighbor grouping (K=2 to 10) (ablation) | CapsNet with simple backbone (M=2) · CIFAR-10 | accuracy: Peak at K=6 (87.4%); K=2 86.6%, K=10 85.2% | Mixed |
Closest papers by meaning
- Capsules with Inverted Dot-Product Attention Routing (ICLR 2020 · similarity 0.89)
- Matrix capsules with EM routing (ICLR 2018 · similarity 0.88)
- Dynamic Routing Between Capsules (NeurIPS 2017 · similarity 0.88)
- Why Capsule Neural Networks Do Not Scale: Challenging the Dynamic Parse-Tree Assumption (AAAI 2023 · 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 0.85)
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.