Research map / Architectures and efficiency / CNN design
Exploiting Cyclic Symmetry in Convolutional 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 |
|---|---|---|---|
| add cyclic slicing at input and mean cyclic pooling before output (ablation) | plankton baseline CNN · Plankton | cross-entropy (test): 0.7035 -> 0.6781 | Helped |
| add cyclic rolling layers after all conv layers and first dense layer (filters reduced 4x) (ablation) | plankton network · Plankton | cross-entropy (test): 0.7035 -> 0.6961 (slight improvement with 0.95M vs 2.84M params) | Mixed |
| add rolling layer after first dense layer with units halved (ablation) | plankton network · Plankton | cross-entropy (test): 0.7035 -> 0.6764 | Helped |
| halve the filters in baseline (parameter-matched control) (ablation) | plankton baseline · Plankton | cross-entropy (test): 0.7035 -> 0.7462 | Hurt |
| add cyclic slicing and mean pooling (ablation) | galaxies baseline CNN · Galaxies | RMSE (test): 0.08343 -> 0.08214 | Helped |
| add rolling layers (filters reduced 4x) (ablation) | galaxies network · Galaxies | RMSE (test): 0.08343 -> 0.08261 with 6.84M params vs 14.57M | Helped |
| add rolling layer after first dense layer with units halved (ablation) | galaxies network · Galaxies | RMSE (test): 0.08343 -> 0.08200 at same 14.57M params | Helped |
| halve filters in baseline (parameter-matched control) (ablation) | galaxies baseline · Galaxies | RMSE (test): 0.08343 -> 0.08710 | Hurt |
| add cyclic slicing and mean pooling (ablation) | Massachusetts buildings baseline · Massachusetts buildings | AUC (test): 0.9714 -> 0.9722 | Helped |
| add rolling layers with filters reduced 4x (ablation) | Massachusetts buildings network · Massachusetts buildings | AUC (test): 0.9714 -> 0.9676 at 158k params | Hurt |
| add rolling layers with filters reduced 2x (ablation) | Massachusetts buildings network · Massachusetts buildings | AUC (test): 0.9714 -> 0.9742 at 598k params (similar to baseline) | Helped |
| choice of pooling function (mean vs max, RMS, with/without ReLU) (ablation) | plankton pooling layer · Plankton | cross-entropy (test): mean without ReLU best (0.6781); max+ReLU worst (0.7085) | Mixed |
Closest papers by meaning
- On Making $SE(2)$-Invariant Networks Optimal (NeurIPS 2026 · similarity 0.90)
- Beyond the Permutation Symmetry of Transformers: The Role of Rotation for Model Fusion (ICML 2025 · similarity 0.89)
- Probing Equivariance and Symmetry Breaking in Convolutional Networks (NeurIPS 2025 · similarity 0.89)
- Relaxed Rotational Equivariance via G-Biases in Vision (AAAI 2025 · similarity 0.88)
- R2Det: Exploring Relaxed Rotation Equivariance in 2D Object Detection (ICLR 2025 · similarity 0.87)
- Bridging Symmetry and Robustness: On the Role of Equivariance in Enhancing Adversarial Robustness (NeurIPS 2025 · similarity 0.87)
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.