Research map / Learning theory and algorithms / Combinatorial and search optimization
Design Linear Constrained Neural Layers with Implicit Convex Optimization
Junchi Yan, Jiaxi Liu, Yihui Tu, Fangyuan Zhou, Wenzheng Pan, Zhongteng Gui, Liangliang Shi
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 Sinkhorn with BLCLayer-10 iters | NGMv2 · Pascal VOC Keypoint partial graph matching (unfiltered) | F1: 0.5453 -> 0.6012 | Improved |
| replace LinSAT-5 with BLCLayer-5 iters | NGMv2 · Pascal VOC Keypoint partial graph matching (unfiltered) | F1: 0.5858 -> 0.5972 | Improved |
| replace LinSAT-5 with BLCLayer-5 iters | NGMv2 · Pascal VOC Keypoint partial graph matching (unfiltered) | training time: 1h 56m -> 1h 28m | Improved |
| raise temperature tau from 0.01 to 0.1 in LinSAT (ablation) | NGMv2 · Pascal VOC Keypoint partial graph matching (unfiltered) | F1: 0.5839 -> 0.5041 | Hurt |
| raise temperature tau from 0.01 to 0.1 in BLCNet (ablation) | NGMv2 · Pascal VOC Keypoint partial graph matching (unfiltered) | F1: 0.6012 -> 0.5921 | Hurt |
| increase LinSAT iterations from 10 to 40 (ablation) | NGMv2 · Pascal VOC Keypoint partial graph matching (unfiltered) | F1: 0.5839 -> 0.5894 | No clear effect |
| use GLCNet instead of LinSAT-100 | StemGNN · Portfolio allocation (S&P 500) | Sharpe ratio: 2.755 -> 3.053 | Improved |
| use BLCNet-5 instead of LinSAT-100 | StemGNN · Portfolio allocation (S&P 500) | training time: 4862.5 s -> 202.65 s | Improved |
| add linear constraint layer vs unconstrained | StemGNN · Portfolio allocation (S&P 500) | Sharpe ratio: 2.190 -> 3.053 | Improved |
| reduce LinSAT test temperature from 1/200 to 1/10 (ablation) | LinSAT · Portfolio allocation (S&P 500), ablation | Sharpe ratio: 2.124 -> 1.311 | Hurt |
| increase GLCNet iterations from 5 to 10 (ablation) | GLCNet · Portfolio allocation (S&P 500), ablation | Sharpe ratio: 3.000 -> 3.027 | No clear effect |
| use GLCNet instead of HardNet | LP solver · Linear programming (3 mixed constraint scales) | objective percentage: 0.908 -> 0.956 (500 vars); 0.869 -> 0.925 (2k vars) | Improved |
Closest papers by meaning
- Enforcing Hard Linear Constraints in Deep Learning Models with Decision Rules (NeurIPS 2025 · similarity 0.85)
- GLinSAT: The General Linear Satisfiability Neural Network Layer By Accelerated Gradient Descent (NeurIPS 2024 · similarity 0.85)
- On the Power of Small-size Graph Neural Networks for Linear Programming (NeurIPS 2024 · similarity 0.85)
- BPQP: A Differentiable Convex Optimization Framework for Efficient End-to-End Learning (NeurIPS 2024 · similarity 0.84)
- On The Concurrence of Layer-wise Preconditioning Methods and Provable Feature Learning (ICML 2025 · similarity 0.84)
- Learning Initial Basis Selection for Linear Programming via Duality-Inspired Tripartite Graph Representation and Comprehensive Supervision (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.
