Research map / Learning theory and algorithms / Optimal transport
Re-Basin via Implicit Sinkhorn Differentiation
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 Sinkhorn re-basin with costRnd instead of STE | Linear mode connectivity · First degree polynomial | AUC: 0.11$\pm$0.10 -> 0.01$\pm$0.01 | Helped |
| Use Sinkhorn re-basin with costRnd instead of STE | Linear mode connectivity · Third degree polynomial | Barrier: 0.24$\pm$0.23 -> 0.03$\pm$0.03 | Helped |
| Use Sinkhorn re-basin with costRnd instead of WM | Linear mode connectivity · Mnist | AUC: 0.01$\pm$0.00 -> 0.00$\pm$0.00 | No clear effect |
| Use Sinkhorn re-basin with costMid instead of STE | Linear mode connectivity · Cifar10 | Barrier: 0.15$\pm$0.02 -> 0.05$\pm$0.01 | Helped |
| Use Sinkhorn Re-basin with costL2 instead of WM | Linear mode connectivity · First degree polynomial | AUC: 0.16$\pm$0.15 -> 0.05$\pm$0.06 | Helped |
| Use Rebasin with replay instead of A-GEM | Continual learning · Rotated Mnist | Accuracy: 68.47$\pm$0.90 -> 78.14$\pm$0.50 | Helped |
| Use Rebasin with replay instead of A-GEM | Continual learning · Rotated Mnist | Forgetting: 0.28$\pm$0.01 -> 0.12$\pm$0.01 | Helped |
| Use Rebasin with replay instead of EWC | Continual learning · Split Cifar100 | Accuracy: 50.50$\pm$1.33 -> 51.34$\pm$0.74 | Helped |
| Use Rebasin with replay instead of EWC | Continual learning · Split Cifar100 | Forgetting: 0.24$\pm$0.02 -> 0.07$\pm$0.02 | Helped |
| Replace WM with Sinkhorn re-basin using squared L2 matching | Optimal transport matching · Random initialization, 2 hidden layers | L1 distance (scaled × 10^3): 6.05$\pm$9.17 -> 0.00$\pm$0.00 | Helped |
| Replace WM with Sinkhorn re-basin using squared L2 matching | Optimal transport matching · Random initialization, 8 hidden layers | L1 distance (scaled × 10^3): 0.50$\pm$1.55 -> 0.00$\pm$0.00 | Helped |
| Use costRnd instead of costMid (ablation) | Sinkhorn re-basin for linear mode connectivity · First degree polynomial | AUC: 0.03$\pm$0.02 -> 0.01$\pm$0.01 | Helped |
Closest papers by meaning
- Git Re-Basin: Merging Models modulo Permutation Symmetries (ICLR 2023 · similarity 0.89)
- On Partial Optimal Transport: Revising the Infeasibility of Sinkhorn and Efficient Gradient Methods (AAAI 2024 · similarity 0.84)
- Accelerating Sinkhorn algorithm with sparse Newton iterations (ICLR 2024 · similarity 0.83)
- RepAn: Enhanced Annealing through Re-parameterization (CVPR 2024 · similarity 0.83)
- On Fitting Flow Models with Large Sinkhorn Couplings (NeurIPS 2026 · similarity 0.83)
- Update Your Transformer to the Latest Release: Re-Basin of Task Vectors (ICML 2025 · similarity 0.82)
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.