Research map / Learning theory and algorithms / Prediction and decision losses
Neural Bregman Divergences for Distance Learning
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 |
|---|---|---|---|
| learn a smooth convex generating function with an ICNN instead of using max-affine components | Bregman divergence learning · Exponential, Gaussian, and multinomial distribution clustering | Purity: 0.665 -> 0.735; 0.867 -> 0.913; 0.876 -> 0.921 | Helped |
| learn a flexible Bregman divergence rather than use a fixed Euclidean distance | NBD · Exponential distribution clustering | Purity: 0.365 -> 0.735 | Helped |
| use a learned Bregman divergence rather than Mahalanobis distance | NBD · Asymmetric divergence regression with x log x targets | MAE: 1.45 -> 0.52 | Helped |
| use NBD rather than Widenorm | deep metric learning · STL10 | MAP@10: 95.1 -> 95.0 | Hurt |
| replace the metric with a learned NBD instead of fixing Euclidean distance | deep metric learning · CIFAR10, STL10, and SVHN | MAP@10: CIFAR10: 95.0 -> 95.6; STL10: 95.0 -> 95.0; SVHN: 96.9 -> 96.9 | No clear effect |
| use a Bregman divergence that can violate the triangle inequality instead of a triangle-inequality-preserving metric | NBD · Shortest-path test graphs | MSE: 3d: 33.49 vs 22.44; taxi: 66.27 vs 16.20 | Hurt |
| learn the square root of the Bregman divergence (ablation) | NBD · Shortest path: octagon test | MSE: 25.32 -> 40.18 | Hurt |
| replace the divergence with the Generalized Symmetrized Bregman divergence (ablation) | NBD · Shortest path: octagon test | MSE: 25.32 -> 43.13 | Hurt |
Closest papers by meaning
- Deep Divergence Learning (ICML 2020 · similarity 0.93)
- Learning to Approximate a Bregman Divergence (NeurIPS 2020 · similarity 0.92)
- Learning Bregman Divergences with Application to Robustness (NeurIPS 2024 · similarity 0.87)
- Learning Distances from Data with Normalizing Flows and Score Matching (ICML 2025 · similarity 0.86)
- Connecting Neural Models Latent Geometries with Relative Geodesic Representations (NeurIPS 2025 · similarity 0.86)
- Euclidean Embedding of Data Using Local Distances (NeurIPS 2026 · similarity 0.85)
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.