Research map / Label-efficient and robust learning / Recognition and long-tail
Cherry-Picking Gradients: Learning Low-Rank Embeddings of Visual Data via Differentiable Cross-Approximation
Mikhail Usvyatsov, Anastasia Makarova, Rafael Ballester-Ripoll, Maxim Rakhuba, Andreas Krause, Konrad Schindler
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 CNN encoder before TT decomposition (ablation) | C-Pic · OSIC (32x512x512) | mLLL: -6.91 -> -6.73 | Improved |
| add CNN encoder before TT decomposition (ablation) | C-Pic · BraTS 256^3 | RMSE (days): 83.9 -> 48.2 | Improved |
| use C-Pic low-rank TT bottleneck instead of 3D ResNet-34 | C-Pic · OSIC (32x512x512) | mLLL: -6.86 -> -6.73 | Improved |
| use C-Pic instead of 3D ResNet-34 | C-Pic · BraTS 256^3 | RMSE (days): 48.7 -> 48.2 | No clear effect |
| raise encoding channel depth from 8 to 16 (ablation) | C-Pic · BraTS 256^3 | RMSE (days): 51.1 -> 49.1 | Improved |
| raise encoding channel depth from 8 to 32 (ablation) | C-Pic · BraTS 256^3 | RMSE (days): 51.1 -> 48.2 | Improved |
| raise TT rank r from 10 to 15 at n_dim=8 (ablation) | C-Pic · BraTS 256^3 | RMSE (days): 51.1 -> 51.1 | No clear effect |
| double resolution per axis (8x voxels) with synthetic upsampling | C-Pic · OSIC | mLLL: -6.73 -> -6.73 | No clear effect |
| double resolution per axis with synthetic upsampling | C-Pic · BraTS (n_dim=8) | RMSE (days): 51.1 -> 51.2 | No clear effect |
| raise input resolution to 64x1024x1024 for 3D ResNet-34 | 3D ResNet-34 · OSIC | forward/backward memory (GB): 7.0 -> 57.9 | Hurt |
| raise input resolution to 64x1024x1024 for C-Pic | C-Pic · OSIC | forward/backward memory (GB): 3.5 -> 4.2 | Mixed |
| C-Pic versus ResNet-34 training time | C-Pic · OSIC (32x512x512) | training time (s/epoch): 4650 -> 51480 | Hurt |
Closest papers by meaning
- Coarse-To-Fine Tensor Trains for Compact Visual Representations (ICML 2024 · similarity 0.88)
- Distribution-Aware Tensor Decomposition for Compression of Convolutional Neural Networks (NeurIPS 2025 · similarity 0.87)
- Convolutions and More as Einsum: A Tensor Network Perspective with Advances for Second-Order Methods (NeurIPS 2024 · similarity 0.86)
- Deep Rank-One Tensor Functional Factorization for Multi-Dimensional Data Recovery (AAAI 2025 · similarity 0.86)
- Low-Rank Tensor Transitions (LoRT) for Transferable Tensor Regression (ICML 2025 · similarity 0.86)
- Performance Gaps in Multi-view Clustering under the Nested Matrix-Tensor Model (ICLR 2024 · 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.
