Research map / Learning theory and algorithms / Combinatorial and search optimization
Differentiable Optimization of Generalized Nondecomposable Functions using Linear Programs
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 |
|---|---|---|---|
| optimize binary AUC directly with the LP solver instead of cross-entropy | ResNet with Newton LP layer · CIFAR100, positive ratio 91% | AUC (%): 57.8 -> 58.2 | Helped |
| optimize binary AUC with LP solver instead of cross-entropy | ResNet with Newton LP layer · CIFAR10, positive ratio 91% | AUC (%): 65.2 -> 72.5 | Helped |
| LP-based AUC solver instead of PPD-SG surrogate | ResNet with Newton LP layer · STL10, positive ratio 91% | AUC (%): 70.7 -> 68.4 | Hurt |
| LP-based AUC solver instead of PPD-SG surrogate | ResNet with Newton LP layer · CIFAR100, positive ratio 50% | AUC (%): 65.2 -> 69.0 | Helped |
| LP-based AUC solver instead of PPD-SG on balanced data | ResNet with Newton LP layer · Cat&Dog, positive ratio 50% | AUC (%): 87.1 -> 87.8 | No clear effect |
| optimize multiclass AUCmu with LP solver instead of cross-entropy | ResNet with Newton LP layer · STL10, drop rate 90% | AUCmu (%): 68.9 -> 76.1 | Helped |
| optimize multiclass AUC (one-vs-all) with LP solver instead of cross-entropy | ResNet with Newton LP layer · CIFAR100, drop rate 80% | AUCova (%): 59.4 -> 59.2 | Hurt |
| optimize multiclass AUCmu with LP solver instead of cross-entropy | ResNet with Newton LP layer · CIFAR100, drop rate 0% | AUCmu (%): 64.0 -> 67.2 | Helped |
| optimize F1-score directly with LP solver instead of cross-entropy | ResNet with Newton LP layer · STL10 | accuracy (%): 71.8% -> 72.7% | Helped |
| optimize F1-score directly with LP solver instead of cross-entropy | ResNet with Newton LP layer · CIFAR100 | accuracy (%): 60.4% -> 63.4% | Helped |
| use CVXPY-SCS or AP-Perf in place of the Newton LP solver for F1 optimization (ablation) | F1 optimization · four datasets | F1-score: 66.7% | Hurt |
| epsilon 0.001 instead of 0.01 in AUC LP (ablation) | LP formulation · Cat&Dog | AUC | Hurt |
Closest papers by meaning
- DOGE-Train: Discrete Optimization on GPU with End-to-End Training (AAAI 2024 · similarity 0.87)
- Random Sparse Lifts: Construction, Analysis and Convergence of finite sparse networks (ICLR 2024 · similarity 0.87)
- BPQP: A Differentiable Convex Optimization Framework for Efficient End-to-End Learning (NeurIPS 2024 · similarity 0.87)
- Learning Generalized Linear Programming Value Functions (NeurIPS 2024 · similarity 0.87)
- Stacey: Promoting Stochastic Steepest Descent via Accelerated $\ell_p$-Smooth Nonconvex Optimization (ICML 2025 · similarity 0.86)
- Do Neural Networks Need Gradient Descent to Generalize? A Theoretical Study (NeurIPS 2025 · similarity 0.86)
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.