Research map / Probabilistic and scientific modeling / Molecules and proteins
Equivariant Adaptation of Large Pretrained Models
Arnab Kumar Mondal, Siba Smarak Panigrahi, Sékou-Oumar Kaba, Sai Rajeswar, Siamak Ravanbakhsh
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 prior regularization loss to learned canonicalization (ablation) | ResNet50 (LC) · CIFAR10 test | accuracy: 93.29 -> 96.19 | Improved |
| add prior regularization loss to learned canonicalization (ablation) | ResNet50 (LC) · CIFAR100 test | accuracy: 78.50 -> 83.44 | Improved |
| add prior regularization loss to learned canonicalization (ablation) | ViT (LC) · CIFAR100 C8-averaged | accuracy: 80.48 -> 83.61 | Improved |
| add prior regularization loss to learned canonicalization (ablation) | ResNet50 (LC) · STL10 C8-averaged | accuracy: 93.92 -> 96.37 | Improved |
| use equivariant canonicalization versus vanilla fine-tuning | ResNet50 · CIFAR10 C8-averaged | accuracy: 57.77 -> 95.31 | Improved |
| use equivariant canonicalization versus vanilla fine-tuning | ResNet50 · CIFAR10 test | accuracy: 96.97 -> 96.19 | Hurt |
| rotation augmentation -180 to +180 versus -10 to +10 for training | ResNet50 pretrained · CIFAR100 test | accuracy: 85.84 -> 80.21 | Hurt |
| train from scratch with learned canonicalization instead of small rotation augmentation | ResNet50 (scratch) · CIFAR10 test | accuracy: 90.96 -> 83.11 | Hurt |
| use C8 rotation augmentation instead of full rotation augmentation | ResNet50 · CIFAR100 C8-averaged | accuracy: 74.12 -> 79.72 | Improved |
| make canonicalization network more expressive (G-CNN to G-WRN) (ablation) | MaskRCNN + canonicalization · COCO 2017 val | mask mAP: 35.77 -> 44.51 | Improved |
| make canonicalization network more expressive (G-CNN to G-WRN) (ablation) | SAM + canonicalization · COCO 2017 val | mask mAP: 59.28 -> 62.13 | Improved |
| add prior-regularized canonicalization to MaskRCNN | MaskRCNN · COCO 2017 val C4-averaged | mask mAP: 27.67 -> 44.50 | Improved |
Closest papers by meaning
- Improving Equivariant Model Training via Constraint Relaxation (NeurIPS 2024 · similarity 0.90)
- Learning (Approximately) Equivariant Networks via Constrained Optimization (NeurIPS 2025 · similarity 0.90)
- Approximately Equivariant Neural Processes (NeurIPS 2024 · similarity 0.90)
- Improving Equivariant Networks with Probabilistic Symmetry Breaking (ICLR 2025 · similarity 0.88)
- A Tale of Two Symmetries: Exploring the Loss Landscape of Equivariant Models (NeurIPS 2025 · similarity 0.88)
- Interpreting Equivariant Representations (ICML 2024 · similarity 0.88)
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.
