Research map / Architectures and efficiency / RNNs and normalization
Learning Probabilistic Symmetrization for Architecture Agnostic Equivariance
Jinwoo Kim, Dat Tien Nguyen, Ayhan Suleymanzade, Hyeokjun An, Seunghoon Hong
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 uniform group averaging with a learned probabilistic distribution | MLP symmetrization · EXP-classify | test accuracy: 50% -> 100% | Improved |
| use frame averaging instead of learned probabilistic symmetrization | MLP symmetrization · EXP-classify | test accuracy: 100% -> 100% | No clear effect |
| add stochastic sampling rather than fixing ε at initialization (ablation) | MLP-PS · EXP-classify | test accuracy: 79.5% -> 100% | Improved |
| add probabilistic sampling instead of using a deterministic canonicalizer (ablation) | MLP symmetrization · EXP-classify | test accuracy: 50% -> 100% | Improved |
| learn the symmetrization distribution instead of using uniform group averaging | Transformer symmetrization · n-body | Position MSE: 0.00414 ± 0.00001 -> 0.00401 ± 0.00001 | Improved |
| add probabilistic sampling instead of one deterministic group element per input (ablation) | Transformer symmetrization · n-body | Position MSE: 0.00508 -> 0.00401 ± 0.00001 | Improved |
| add 10-sample symmetrization at test time instead of 1 sample (ablation) | ViT-PS without pretraining · PATTERN | test accuracy: 85.868 ± 0.017 -> 85.989 ± 0.011 | Improved |
| add ImageNet-21k pretraining (ablation) | ViT-PS, 10-sample · PATTERN | test accuracy: 85.989 ± 0.011 -> 86.650 ± 0.010 | Improved |
| add 10-sample symmetrization at test time instead of 1 sample (ablation) | ViT-PS pretrained on ImageNet-21k · PATTERN | test accuracy: 86.573 ± 0.030 -> 86.650 ± 0.010 | Improved |
| use learned probabilistic symmetrization instead of canonicalization | ViT pretrained on ImageNet-21k · PATTERN | test accuracy: 86.534 -> 86.650 ± 0.010 | Improved |
| use ImageNet-21k pretrained ViT with probabilistic symmetrization rather than the strongest listed graph baseline | PATTERN node classification · PATTERN | test accuracy: 86.245 -> 86.650 ± 0.010 | Improved |
| use ViT-PS instead of GraphGPS | Peptides-func · Peptides-func | AP: 0.6535 -> 0.6575 | Improved |
Closest papers by meaning
- Improving Equivariant Networks with Probabilistic Symmetry Breaking (ICLR 2025 · similarity 0.91)
- Learning symmetries via weight-sharing with doubly stochastic tensors (NeurIPS 2024 · similarity 0.91)
- Probing Equivariance and Symmetry Breaking in Convolutional Networks (NeurIPS 2025 · similarity 0.91)
- A Tale of Two Symmetries: Exploring the Loss Landscape of Equivariant Models (NeurIPS 2025 · similarity 0.91)
- On the hardness of learning under symmetries (ICLR 2024 · similarity 0.90)
- Learning (Approximately) Equivariant Networks via Constrained Optimization (NeurIPS 2025 · similarity 0.90)
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.
