Research map / Learning theory and algorithms / Prediction and decision losses
Neural approximation of Wasserstein distance via a universal architecture for symmetric and factorwise group invariant functions
Samantha Chen, Yusu Wang
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 ProductNet instead of WPCE (Siamese autoencoder with Sinkhorn loss) | 1-Wasserstein approximation · noisy-sphere-3, input size [100, 300] | mean relative error: 0.341 ± 0.202 -> 0.046 ± 0.043 | Improved |
| Evaluate on input sizes unseen at training time vs WPCE | ProductNet · ModelNet-small, input size [300, 500] | mean relative error: 0.241 ± 0.198 -> 0.111 ± 0.086 | Improved |
| Use ProductNet instead of WPCE on seen sizes | 1-Wasserstein approximation · ModelNet-small, input size [20, 200] | mean relative error: 0.077 ± 0.075 -> 0.084 ± 0.077 | Hurt |
| Use ProductNet instead of Siamese DeepSets | 1-Wasserstein approximation · uniform, input size [200, 300] | mean relative error: 0.917 ± 0.869 -> 0.131 ± 0.096 | Improved |
| Replace WPCE with ProductNet to cut training cost | ProductNet · RNAseq(2k) | training time: 14hr 26min -> 15min | Improved |
| Replace WPCE with ProductNet to cut training cost | ProductNet · noisy-sphere-3 | epochs to convergence: 100 -> 20 | Improved |
| Compare ProductNet with Sinkhorn (eps=0.1) on high-dimensional unseen-size data | ProductNet · RNAseq, input size [300, 500] | mean relative error: 0.048 ± 0.006 -> 0.292 ± 0.041 | Hurt |
| Use ProductNet instead of Sinkhorn (eps=0.1) on inference speed | ProductNet · ModelNet-large | time for 500 computations (s): 239.448 -> 1.174 | Improved |
| Use ProductNet instead of Sinkhorn (eps=0.1) on accuracy | ProductNet · noisy-sphere-6, input size [100, 300] | mean relative error: 0.137 ± 0.122 -> 0.015 ± 0.014 | Improved |
| Lower Sinkhorn regularization from eps=0.1 to eps=0.01 | Sinkhorn · ModelNet-small | mean relative error and time (s): 0.187 ± 0.232 -> 0.011 ± 0.003; 2.074 -> 104.712 | Mixed |
| Lower Sinkhorn regularization from eps=0.1 to eps=0.01 | Sinkhorn · noisy-sphere-6 | mean relative error: 0.137 ± 0.122 -> 0.326 ± 0.135 | Hurt |
| Use PointNet (max aggregator) instead of DeepSets (sum) as WPCE encoder (ablation) | WPCE · 1-Wasserstein approximation | approximation quality | No clear effect |
Closest papers by meaning
- Augmented Sliced Wasserstein Distances (ICLR 2022 · similarity 0.86)
- Deep Neural Network Approximation of Invariant Functions through Dynamical Systems (JMLR 2024 · similarity 0.86)
- On the Completeness of Invariant Geometric Deep Learning Models (ICLR 2025 · similarity 0.85)
- Siegel Neural Networks (NeurIPS 2025 · similarity 0.84)
- Monomial Matrix Group Equivariant Neural Functional Networks (NeurIPS 2024 · similarity 0.84)
- Global curvature for second-order optimization of neural networks (ICML 2025 · similarity 0.84)
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.
