Research map / Architectures and efficiency / CNN design
Learning Implicitly Recurrent CNNs Through Parameter Sharing
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 soft parameter sharing with one template per layer | WRN 28-10 · CIFAR-10 test | test error (%): 4.0 -> 3.74 | Helped |
| Use soft parameter sharing with one template per layer | WRN 28-10 · CIFAR-100 test | test error (%): 19.25 -> 18.78 | Helped |
| Reduce sharing templates to one per group | WRN 28-10 · CIFAR-10 and CIFAR-100 test | test error (%): 4.0 -> 4.01; 19.25 -> 19.73 | Mixed |
| Reduce sharing templates to 2 per group | SWRN 28-18* · CIFAR-10 test | test error (%): 3.48 -> 3.43 | Helped |
| Reduce sharing templates to 2 per group | SWRN 28-18* · CIFAR-100 test | test error (%): 17.43 -> 17.75 | Hurt |
| Apply soft parameter sharing without reducing parameter count | WRN 50-2 · ImageNet validation | Top-1 error (%): 22.0 -> 21.74 | Helped |
| Apply soft parameter sharing without reducing parameter count | WRN 50-2 · ImageNet validation | Top-5 error (%): 6.05 -> 5.95 | Helped |
| Use soft parameter sharing with one template per layer | CNN · Shortest paths curriculum | F1 score | Helped |
| Add recurrence regularizer λ_R = 0.01 | SCNN · Shortest paths curriculum | F1 score | Mixed |
| Fold six stage-2 layers into a two-layer loop (ablation) | SWRN 28-10-4 · CIFAR-10 test | test error increase (%): 0.02% | Hurt |
| Initialize coefficient matrix orthogonally instead of uniformly or normally (ablation) | Soft-sharing CNN · Classification experiments | performance | Helped |
| Initialize coefficient matrix as identity when L=k (ablation) | Soft-sharing CNN · Classification experiments | performance | No clear effect |
Closest papers by meaning
- Recurrent Convolutional Neural Networks Learn Succinct Learning Algorithms (NeurIPS 2022 · similarity 0.91)
- Task-Driven Convolutional Recurrent Models of the Visual System (NeurIPS 2018 · similarity 0.89)
- Pixel Recurrent Neural Networks (ICML 2016 · similarity 0.89)
- Deeply Shared Filter Bases for Parameter-Efficient Convolutional Neural Networks (NeurIPS 2021 · similarity 0.88)
- Recurrent neural network dynamical systems for biological vision (NeurIPS 2024 · similarity 0.88)
- RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks (ICLR 2025 · 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.