Research map / Architectures and efficiency / Pruning and efficient tuning
Neural Network Encapsulation
Hongyang Li, Xiaoyang Guo, Bo DaiWanli Ouyang, Xiaogang Wang
Read the paperCheck 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 |
|---|---|---|---|
| replace master-only capConv with master/aide approximate routing (ablation) | EncapNet 6-layer · CIFAR-10 (6-layer) | error rate (%): 13.87 -> 11.93 | Improved |
| replace dynamic-routing CapNet with EncapNet master/aide (parameters) | 6-layer network · CIFAR-10 (6-layer) | param count (Mb): 151.24 -> 60.68 | Improved |
| replace dynamic-routing CapNet with EncapNet master/aide | 6-layer network · CIFAR-10 (6-layer) | error rate (%): 14.28 -> 11.93 | Improved |
| replace EM-routing CapNet with EncapNet master/aide | 6-layer network · MNIST (6-layer) | error rate (%): 0.31 -> 0.25 | Improved |
| add aide branch (master/aide v3) to master-only capConv (ablation) | EncapNet-18 · CIFAR-10 (18-layer ablation, capConv design) | error rate (%): 9.83 -> 7.41 | Improved |
| add skip connections on both types to master/aide v3 (ablation) | EncapNet-18 · CIFAR-10 (18-layer ablation, capConv design) | error rate (%): 7.41 -> 5.82 | Improved |
| add a second OT divergence to each module (ablation) | EncapNet-18 · CIFAR-10 (18-layer ablation, capConv design) | error rate (%): 5.82 -> 4.55 | Improved |
| add OT (Sinkhorn) feedback loss instead of KL loss (ablation) | EncapNet-18 · CIFAR-10 (18-layer ablation, regularization) | error rate (%): 5.31 -> 4.58 | Improved |
| add OT feedback loss to ResNet-18 (ablation) | ResNet-18 · CIFAR-10 (18-layer ablation, regularization) | error rate (%): 8.03 -> 7.67 | Improved |
| add bias fix to Sinkhorn divergence (ablation) | EncapNet-18 OT loss · CIFAR-10 (18-layer ablation, regularization) | error rate (%): 4.71 -> 4.58 | Improved |
| add feature extractor f_psi before comparing inputs (ablation) | EncapNet-18 OT loss · CIFAR-10 (18-layer ablation, regularization) | error rate (%): 5.79 -> 4.58 | Improved |
| use capsule generator instead of vanilla CNN generator g_phi (ablation) | EncapNet-18 OT loss · CIFAR-10 (18-layer ablation, regularization) | error rate (%): 5.01 -> 4.58 | Improved |
Closest papers by meaning
- Dynamic Routing Between Capsules (NeurIPS 2017 · similarity 0.89)
- Self-Routing Capsule Networks (NeurIPS 2019 · similarity 0.88)
- Ladder Capsule Network (ICML 2019 · similarity 0.88)
- Matrix capsules with EM routing (ICLR 2018 · 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)
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.
