Research map / Learning theory and algorithms / Optimal transport
Efficient Sliced Wasserstein Distance Computation via Adaptive Bayesian Optimization
Manish Acharya, David Hyde
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 one-shot BO-learned directions (BOSW) instead of QSW/MC for approximating SW | BOSW · ShapeNet point-cloud SW approximation error | absolute error | Hurt |
| periodically refresh BO directions (RBOSW) during gradient flow (ablation) | BOSW · Point-cloud interpolation (L=100) | W2 at step 100: 6.096 -> 4.155 (at step 100) | Improved |
| periodically refresh BO directions (RBOSW) increases runtime (ablation) | BOSW · Point-cloud interpolation (L=100) | runtime (s): 3.99 -> 44.45 | Hurt |
| seed BO from QSW and restart periodically (ARBOSW) instead of one-shot ABOSW (ablation) | ABOSW · Point-cloud interpolation (L=100) | W2 at step 500: 0.076 -> 0.003 | Improved |
| add QSW seeding with a few BO refinements (ABOSW) to BOSW (ablation) | BOSW · Deep point-cloud autoencoder | W2 at epoch 100: 10.34 -> 10.27 | Improved |
| use restart variants (RBOSW, ARBOSW) instead of non-restart BOSW/ABOSW (ablation) | BO selectors · Deep point-cloud autoencoder | SW2 at epoch 400: ABOSW 1.81 vs ARBOSW 1.85; BOSW 1.80 vs RBOSW 1.90 | Hurt |
| use BO selectors instead of recommended CQSW | ABOSW · Deep point-cloud autoencoder | SW2 at epoch 400: 1.84 -> 1.81 | Improved |
| use BO versus QSW on synthetic fitness landscapes | BO projection selector · Quadratic landscape | best fitness: 2.69 vs <= 1.0 at L=10 | Improved |
| use UCB acquisition instead of EI, Thompson sampling or log-EI (ablation) | BO selectors · Point-cloud interpolation | convergence | Improved |
| set UCB beta to 0.7 versus other values (ablation) | BO selectors · Point-cloud interpolation | convergence | Improved |
| set cosine-similarity cutoff to 0.98 versus 0.9 or 0.95 (ablation) | BO selectors · Point-cloud interpolation (L=100) | final error | Improved |
| vary candidate pool size n_c from 512 to 16384 (ablation) | BO selectors · Point-cloud interpolation | error after 500 steps | No clear effect |
Closest papers by meaning
- Quasi-Monte Carlo for 3D Sliced Wasserstein (ICLR 2024 · similarity 0.92)
- Tree-Sliced Wasserstein Distance with Nonlinear Projection (ICML 2025 · similarity 0.90)
- Towards Understanding Gradient Dynamics of the Sliced-Wasserstein Distance via Critical Point Analysis (ICML 2025 · similarity 0.89)
- Differentiable Generalized Sliced Wasserstein Plans (NeurIPS 2025 · similarity 0.89)
- Distance-Based Tree-Sliced Wasserstein Distance (ICLR 2025 · similarity 0.88)
- Sliced Wasserstein with Random-Path Projecting Directions (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.
