Research map / Learning theory and algorithms / Optimal transport
Importance Sparsification for Sinkhorn Algorithm
Mengyu Li, Jun Yu, Tao Li, Cheng Meng
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 |
|---|---|---|---|
| use Spar-Sink instead of Rand-Sink at s=s0(n) | Sinkhorn subsampling · Echocardiogram ED prediction, original scale | ED prediction error: 0.21 -> 0.09 | Improved |
| use Spar-Sink instead of Nys-Sink at s=8s0(n) | Sinkhorn subsampling · Echocardiogram ED prediction, original scale | ED prediction error: 0.32 -> 0.06 | Improved |
| use Spar-Sink instead of full Sinkhorn | WFR distance computation · Echocardiogram ED prediction, original scale | CPU time (s): 15649.31 -> 357.46 | Improved |
| increase subsample size from s0(n) to 8s0(n) for Spar-Sink (ablation) | Spar-Sink · Echocardiogram ED prediction, original scale | ED prediction error: 0.09 -> 0.06 | Improved |
| mean-pool images from 112x112 to 56x56 before Sinkhorn | Sinkhorn · Echocardiogram ED prediction | ED prediction error: 0.06 -> 0.11 | Hurt |
| mean-pool images before Sinkhorn | Sinkhorn · Echocardiogram ED prediction | CPU time (s): 15649.31 -> 668.81 | Improved |
| use Spar-Sink on original images instead of Sinkhorn on pooled images | WFR distance computation · Echocardiogram ED prediction | ED prediction error: 0.11 -> 0.06 | Improved |
| use Spar-Sink on pooled images instead of Rand-Sink at s=s0(n) | Sinkhorn subsampling · Echocardiogram ED prediction, pooled | ED prediction error: 0.38 -> 0.30 | Improved |
| use sparsified Sinkhorn in the autoencoder (SSAE) instead of Sinkhorn (SAE) | Sinkhorn autoencoder · Image generator learning (digits) | FID: 24.72 -> 23.65 | Improved |
| use SSAE instead of SAE | Sinkhorn autoencoder · Image generator learning (digits) | time per epoch (s): 125.87 -> 64.81 | Improved |
| decrease epsilon in Sinkhorn (ablation) | classical Sinkhorn · Synthetic OT, CPU time | CPU time | Hurt |
| use Spar-Sink instead of Rand-Sink and Nys-Sink on synthetic OT and UOT | Sinkhorn subsampling · Synthetic data C1-C3 | RMAE | Improved |
Closest papers by meaning
- Accelerating Sinkhorn algorithm with sparse Newton iterations (ICLR 2024 · similarity 0.87)
- Safe and Sparse Newton Method for Entropic-Regularized Optimal Transport (NeurIPS 2024 · similarity 0.87)
- Debiased Sinkhorn barycenters (ICML 2020 · similarity 0.85)
- The Sparse-Plus-Low-Rank Quasi-Newton Method for Entropic-Regularized Optimal Transport (ICML 2025 · similarity 0.84)
- On Partial Optimal Transport: Revising the Infeasibility of Sinkhorn and Efficient Gradient Methods (AAAI 2024 · similarity 0.84)
- Sparsistency for inverse optimal transport (ICLR 2024 · similarity 0.83)
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.
