Research map / Learning theory and algorithms / Optimal transport
Geometry-Misalignment in Distributional Learning
Tao Wang, Xiaoting Zhong
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 geometry-aware preconditioned updates instead of Euclidean SGD | Sinkhorn distribution matching · DGP-1 high misalignment | iterations to converge: 1600 -> 130 | Improved |
| add misalignment gating (GCO) to always-on geometry-aware updates (ablation) | geometry-aware optimization · DGP-1 high misalignment | wall-clock time: 3.30 -> 1.70 (iterations 130 -> 145) | Improved |
| add gating to GCO (versus no-gate variant) (ablation) | GCO · DGP-1 low misalignment | wall-clock time: 2.60 -> 1.50 | Improved |
| add gating to GCO (ablation) | GCO · DGP-1 high misalignment | iterations / objective gap: iterations 135 -> 145; gap 7.1e-5 -> 7.5e-5 | Hurt |
| use Shampoo or K-FAC instead of SGD | Euclidean first-order baseline · DGP-1 high misalignment | iterations and overhead: iterations 1600 -> 720 (Shampoo), 690 (K-FAC); overhead 1.27 and 1.43 vs 1.00 | Mixed |
| use Adam or AdaGrad instead of SGD | Euclidean first-order baseline · DGP-1 high misalignment | iterations: 1600 -> 1450 (Adam), 1700 (AdaGrad) | No clear effect |
| increase geometry-misalignment from low to high (ablation) | SGD · DGP-1 Linear Anisotropy | iterations to converge: 180 -> 1600 | Hurt |
| increase misalignment from low to high (ablation) | GCO · DGP-1 Linear Anisotropy | fraction of geometry-aware updates (%): 6 -> 34 | Improved |
| use GCO instead of domain adaptation baselines (CORAL, DANN, CDAN, MMD) | distribution matching · DGP-1 high misalignment | objective gap: 1.1 × 10−3 (CDAN) -> 7.5 × 10−5 | Improved |
Closest papers by meaning
- Geometric Losses for Distributional Learning (ICML 2019 · similarity 0.87)
- Optimizing Distributional Geometry Alignment with Optimal Transport for Generative Dataset Distillation (NeurIPS 2025 · similarity 0.86)
- Direct Distributional Optimization for Provable Alignment of Diffusion Models (ICLR 2025 · similarity 0.86)
- Diffusion Model's Generalization Can Be Characterized by Inductive Biases toward a Data-Dependent Ridge Manifold (NeurIPS 2026 · similarity 0.85)
- No Wrong Turns: The Simple Geometry Of Neural Networks Optimization Paths (ICML 2024 · similarity 0.85)
- Supervised Distributional Reduction via Optimal Transport and Dependence Maximization (NeurIPS 2026 · 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.
