Research map / Detection and segmentation / Open-set and few-shot detection
Stacked Capsule Autoencoders
Adam Kosiorek, Sara Sabour, Yee Whye Teh, Geoffrey E. Hinton
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 |
|---|---|---|---|
| Use object-to-part geometric relationships to explain point constellations | CCAE · constellation point data | segmentation error: 26% -> 2.8% | Improved |
| Train a linear classifier on object capsule presence vectors | SCAE · SVHN | unsupervised classification accuracy: 55.33 (3.4) -> 67.27 (4.5) | Improved |
| Remove posterior sparsity constraints (ablation) | SCAE · MNIST | unsupervised classification accuracy: 95.3 (4.65) -> 97.5 (1.55) | No clear effect |
| Remove prior sparsity (ablation) | SCAE · 40x40 MNIST | unsupervised classification accuracy: 98.7 (0.35) -> 88.2 (6.98) | Hurt |
| Remove both prior and posterior sparsity (ablation) | SCAE · 40x40 MNIST | unsupervised classification accuracy: 98.7 (0.35) -> 82.0 (5.46) | Hurt |
| Remove noise from object-capsule presence logits (ablation) | SCAE · 40x40 MNIST | unsupervised classification accuracy: 98.7 (0.35) -> 98.5 (0.12) | No clear effect |
| Remove noise from all capsule presence logits (ablation) | SCAE · 40x40 MNIST | unsupervised classification accuracy: 98.7 (0.35) -> 78.5 (22.69) | Hurt |
| Remove noise from part-capsule presence logits (ablation) | SCAE · 40x40 MNIST | unsupervised classification accuracy: 98.7 (0.35) -> 82.8 (24.83) | Hurt |
| Replace affine transformations with similarity transformations (ablation) | SCAE · 40x40 MNIST | unsupervised classification accuracy: 98.7 (0.35) -> 95.9 (1.59) | Hurt |
| Remove object-dependent deformations (ablation) | SCAE · 40x40 MNIST | unsupervised classification accuracy: 98.7 (0.35) -> 87.2 (18.54) | Hurt |
| Replace attention-based part encoder with a linear part encoder (ablation) | SCAE · AFFNIST | accuracy: 92.2 (0.59) -> 50.8 (26.46) | Hurt |
| Replace attention-based part encoder with a convolutional part encoder (ablation) | SCAE · 40x40 MNIST | unsupervised classification accuracy: 98.7 (0.35) -> 97.8 (.98) | No clear effect |
Closest papers by meaning
- Unsupervised Part Representation by Flow Capsules (ICML 2021 · similarity 0.87)
- 3DPointCaps++: Learning 3D Representations with Capsule Networks (IJCV 2022 · similarity 0.86)
- HP-Capsule: Unsupervised Face Part Discovery by Hierarchical Parsing Capsule Network (CVPR 2022 · similarity 0.85)
- Generative 3D Part Assembly via Part-Whole-Hierarchy Message Passing (CVPR 2024 · similarity 0.84)
- Recurrent Complex-Weighted Autoencoders for Unsupervised Object Discovery (NeurIPS 2024 · similarity 0.83)
- Interpreting CLIP with Hierarchical Sparse Autoencoders (ICML 2025 · similarity 0.83)
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.
