Research map / Learning theory and algorithms / Combinatorial and search optimization
Pinet: Optimizing hard-constrained neural networks with orthogonal projection layers
Panagiotis D. Grontas, Antonio Terpin, Efe C. Balta, Raffaello D'Andrea, John Lygeros
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 |
|---|---|---|---|
| replace cvxpylayers projection with Pinet (GPU) | projection layer · Non-convex small | batch inference time [s]: 2.5917 -> 0.0135 | Improved |
| replace cvxpylayers projection with Pinet (CPU) | projection layer · Non-convex small | single inference time [s]: 0.0120 -> 0.0052 | Improved |
| use Pinet instead of cvxpylayers for training | training budget 25 epochs · Non-convex small | training wall-clock: roughly 10 minutes -> 4 and 28 seconds on the GPU and CPU | Improved |
| use Pinet instead of JAXopt implicit layer | projection layer · Non-convex large | median batch inference time [s]: 19.430 -> 0.2804 | Improved |
| use Pinet instead of JAXopt implicit layer | projection layer · Non-convex large | training time: roughly 14 hours for JAXopt | Improved |
| use Pinet batch GPU solve instead of IPOPT solver | Pinet · Non-convex large | median batch inference time [s]: 10720 -> 0.2804 | Improved |
| use Pinet instead of OSQP solver | Pinet · Convex large | median batch inference time [s]: 680.59 -> 0.2800 | Improved |
| OSQP parametric mode (reuse calculations) (ablation) | OSQP solver · Convex large | median batch inference time [s]: 680.59 -> 62.022 | Improved |
| Pinet instead of DC3 on single inference | Pinet · Non-convex small | median single inference time [s]: 0.0019 -> 0.0056 | Hurt |
| increase DC3 correction steps from 10 to 50 (ablation) | DC3 · Non-convex small | median single inference time [s]: 0.0019 -> 0.0080 | Hurt |
| increase DC3 correction steps from 10 to 50 (ablation) | DC3 · Non-convex large | median batch inference time [s]: 0.0248 -> 0.1177 | Hurt |
| decrease DC3 soft penalty from 10.0 to 2.0 and increase correction steps from 10 to 50 (ablation) | DC3 · Non-convex benchmarks | RS and CV | No clear effect |
Closest papers by meaning
- Learning-Augmented Scalable Linear Assignment Problem Optimization via Neural Dual Warm-Starts (ICML 2026 · similarity 0.86)
- FSNet: Feasibility-Seeking Neural Network for Constrained Optimization with Guarantees (NeurIPS 2025 · similarity 0.86)
- Taking the Road Less Scheduled with Adaptive Polyak Steps (NeurIPS 2026 · similarity 0.85)
- SPACE: Unifying Symmetric and Asymmetric Routing Problems for Generalist Neural Solver (NeurIPS 2026 · similarity 0.85)
- Efficient Bisection Projection to Ensure Neural-Network Solution Feasibility for Optimization over General Set (ICML 2025 · similarity 0.85)
- DOGE-Train: Discrete Optimization on GPU with End-to-End Training (AAAI 2024 · 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.
