Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-supervised learning with rotation-invariant kernels
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 a truncated L=3 rotation-invariant kernel instead of SimCLR, AUH, or VICReg regularization | SFRIK · IN20% linear probing, q=8192 | Top-1 accuracy: 47.5 vs 46.1; 46.8; 46.0 | Helped |
| Increase truncation order from L=1 to L=2 (ablation) | SFRIK · IN20% linear probing, q=8192 | Top-1 accuracy: 27.7 -> 47.0 | Helped |
| Increase truncation order from L=2 to L=3 (ablation) | SFRIK · IN20% linear probing, q=8192 | Top-1 accuracy: 47.0 -> 47.5 | Helped |
| Use a truncated L=3 kernel rather than an RBF kernel (ablation) | SFRIK uniformity term · IN20% linear probing, q=8192 | Top-1 accuracy: 41.3 -> 47.5 | Helped |
| Use a truncated L=3 kernel rather than a generalized distance kernel (ablation) | SFRIK uniformity term · IN20% linear probing, q=8192 | Top-1 accuracy: 27.8 -> 47.5 | Helped |
| Increase embedding dimension from q=8192 to q=16384 (ablation) | SFRIK L=2 · IN100% linear classification | Top-1 accuracy: 70.1 -> 70.3 | Helped |
| Increase embedding dimension from q=16384 to q=32768 (ablation) | SFRIK L=2 · IN100% linear classification | Top-1 accuracy: 70.3 -> 70.3 | No clear effect |
| Use L=3 rather than L=2 at q=8192 (ablation) | SFRIK · IN100% linear classification | Top-1 accuracy: 70.1 -> 70.2 | Helped |
| Use SFRIK instead of VICReg | ResNet-50 pretraining · IN100% semi-supervised, 1% labels | Top-1 accuracy: 49.4 -> 46.9 | Hurt |
| Use SFRIK instead of VICReg | ResNet-50 pretraining · IN100% linear classification | Top-1 accuracy: 70.0 -> 70.3 | Helped |
| Use SFRIK rather than VICReg at embedding dimension 16384 | ResNet-50 pretraining · IN100% pretraining, q=16384 | Pretraining wall time: 19% faster | Helped |
| Use SFRIK rather than VICReg at q=32768 and batch size 1024 | ResNet-50 pretraining · IN100% pretraining | Peak memory per GPU: 31.1GB -> 18.8GB | Helped |
Closest papers by meaning
- Non-parametric Representation Learning with Kernels (AAAI 2024 · similarity 0.87)
- Preventing Dimensional Collapse in Self-Supervised Learning via Orthogonality Regularization (NeurIPS 2024 · similarity 0.86)
- SPHERE-JEPA: Spherical Prediction with Homogeneous Embeddings (NeurIPS 2026 · similarity 0.86)
- LDReg: Local Dimensionality Regularized Self-Supervised Learning (ICLR 2024 · similarity 0.85)
- Random features for Grassmannian kernel approximation with bounded rank-one projections (NeurIPS 2026 · similarity 0.85)
- Self-Supervised Representation Learning with Meta Comprehensive Regularization (AAAI 2024 · similarity 0.85)
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.