Research map / Detection and segmentation / CNN features
Tunable Soft Equivariance with Guarantees
Md Ashiqur Rahman, Lim Jun Hao, Jeremiah Jiang, Teck-Yian Lim, Raymond A. Yeh
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 |
|---|---|---|---|
| add soft equivariance projection to ViT | ViT · ImageNet-1K | top-1 accuracy: 81.67 -> 82.28 | Improved |
| add soft equivariance projection to DINOv2 | DINOv2 · ImageNet-1K | top-1 accuracy: 84.27 -> 85.31 | Improved |
| add soft equivariance projection to ResNet-50 | ResNet-50 · ImageNet-1K | invariance error (x10^-2): 00.24 -> 00.11 | Improved |
| add soft equivariance projection to ResNet-50 | ResNet-50 · CIFAR-10 | invariance error (x10^-2): 27.29 -> 24.20 | Improved |
| use canonicalizer instead of soft equivariance | ViT · ImageNet-1K | top-1 accuracy: 81.67 -> 76.51 | Hurt |
| use canonicalizer instead of soft equivariance | ResNet-50 · CIFAR-100 | top-1 accuracy: 83.10 -> 72.90 | Hurt |
| add soft equivariance projection to DINOv2 for segmentation | DINOv2 · PASCAL VOC | mIoU: 89.57 -> 89.48 | No clear effect |
| add soft equivariance projection to ViT for segmentation | ViT · PASCAL VOC | equivariance error (x10^-2): 12.31 -> 11.12 | Improved |
| use smooth cut-off instead of hard threshold in projection (ablation) | ViT · PASCAL VOC | mIoU: 73.92 -> 74.78 | Improved |
| use smooth cut-off instead of hard threshold in projection (ablation) | ViT · PASCAL VOC | equivariance error (x10^-2): 11.74 -> 11.12 | Improved |
| use Schur decomposition instead of SVD to build projection (ablation) | projection operator · 14x14 input | runtime (s): 8.9 · 10^2 -> 2.5 · 10^-1 | Improved |
| add soft equivariance projection to autoregressive transformer | Autoregressive transformer · ZARA1 trajectory | equivariance error (x10^-2): 1.14 -> 0.39 | Improved |
Closest papers by meaning
- A Probabilistic Approach to Learning the Degree of Equivariance in Steerable CNNs (ICML 2024 · similarity 0.90)
- Relaxing Equivariance Constraints with Non-stationary Continuous Filters (NeurIPS 2022 · similarity 0.90)
- Soft Equivariance Regularization for Invariant Self-Supervised Learning (ICLR 2026 · similarity 0.90)
- Equivariant Self-Supervised Learning: Encouraging Equivariance in Representations (ICLR 2022 · similarity 0.88)
- Probing Equivariance and Symmetry Breaking in Convolutional Networks (NeurIPS 2025 · similarity 0.87)
- Learning (Approximately) Equivariant Networks via Constrained Optimization (NeurIPS 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.
