Research map / Learning theory and algorithms / Prediction and decision losses
Learning with Differentiable Pertubed Optimizers
Quentin Berthet, Mathieu Blondel, Olivier Teboul, Marco Cuturi, Jean-Philippe Vert, Francis R. Bach
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 perturbed Fenchel-Young loss instead of blackbox loss | affine model · Label ranking, 21 datasets | fraction of datasets better or similar (within 5%): 76% better, 90% better or similar | Improved |
| use perturbed Fenchel-Young loss instead of blackbox loss | affine model · Label ranking: authorship | Spearman correlation: 0.76 -> 0.95 | Improved |
| use Fenchel-Young loss instead of squared loss | affine model · Label ranking: housing | Spearman correlation: 0.70 -> 0.75 | Improved |
| use Fenchel-Young loss instead of squared loss | affine model · Label ranking: authorship | Spearman correlation: 0.96 -> 0.95 | No clear effect |
| use Fenchel-Young loss instead of squared loss on perturbed argmax (ablation) | affine model · Label ranking: authorship | Spearman correlation: 0.28 -> 0.95 | Improved |
| also tune temperature epsilon for perturbed+squared loss (ablation) | affine model · Label ranking: authorship | Spearman correlation: 0.28 -> 0.93 | Improved |
| very high or very low temperature epsilon (ablation) | perturbed FY · CIFAR-10 vanilla-CNN | train/test squared loss | Hurt |
| perturbed FY loss vs cross entropy | vanilla CNN · CIFAR-10 | test accuracy | No clear effect |
| perturbed FY with single perturbation M=1 vs blackbox/squared | ResNet18 first 5 layers · Warcraft shortest path | perfect accuracy / cost ratio | Improved |
Closest papers by meaning
- Branch & Learn for Recursively and Iteratively Solvable Problems in Predict+Optimize (NeurIPS 2022 · similarity 0.87)
- Accelerating Optimization via Differentiable Stopping Time (NeurIPS 2025 · similarity 0.85)
- Solver-Free Decision-Focused Learning for Linear Optimization Problems (NeurIPS 2025 · similarity 0.85)
- Hybrid Decentralized Optimization: Leveraging Both First- and Zeroth-Order Optimizers for Faster Convergence (AAAI 2025 · similarity 0.85)
- Overcoming the Optimizer's Curse: Obtaining Realistic Prescriptions from Neural Networks (ICML 2024 · similarity 0.84)
- A Generalization Result for Convergence in Learning-to-Optimize (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.
