Research map / Optimization / Optimizers
Injecting Logical Constraints into Neural Networks via Straight-Through Estimators
Zhun Yang, Joohyung Lee, Chiyoun Park
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 |
|---|---|---|---|
| Train with CL-STE constraints rather than NeurASP | CNN · Sudoku (Park 70k, inference trick) | whole-board accuracy: 66.5% -> 93.7% | Improved |
| Replace NeurASP with CL-STE constraint loss | CNN · Sudoku (Park 70k) | training time: 13230 -> 813 minutes | Improved |
| Train with CL-STE instead of SATNet | Sudoku model · Sudoku (Park 70k) | test accuracy: 67.78% -> 93.70% | Improved |
| Train with CL-STE instead of SATNet | Sudoku model · Sudoku (Palm) | test accuracy: 6.76% -> 27.37% | Improved |
| Use CL-STE rather than DeepProbLog | MNIST addition classifier · mnistAdd | training time: 2565s -> 22s | Improved |
| Use CL-STE rather than NeurASP | MNIST addition classifier · mnistAdd2 | training time: 1682s -> 92s | Improved |
| Use CL-STE rather than symbolic baselines | MNIST addition classifier · mnistAdd3 | training completion: timeout -> 402s | Improved |
| Use probability binarization b_p(x) instead of b(x) (ablation) | CL-STE with iSTE · member(5) | accuracy (%): 51.1$\pm$24.9 -> 95.0$\pm$0.5 | Improved |
| Use probability binarization b_p(x) instead of b(x) (ablation) | CL-STE with iSTE · member(3) | accuracy (%): 73.2$\pm$9.1 -> 95.5$\pm$0.7 | Improved |
| Use saturated STE instead of iSTE with b(x) (ablation) | CL-STE · member(5) | accuracy (%): 51.1$\pm$24.9 -> 88.0$\pm$7.1 | Improved |
| Use saturated STE instead of iSTE with b(x) (ablation) | CL-STE · member(3) | accuracy (%): 73.2$\pm$9.1 -> 83.2$\pm$8.4 | Improved |
| Add constraint losses to RRN training while keeping labeled data fixed | RRN · Sudoku (Palm) | test accuracy | Improved |
Closest papers by meaning
- Constraints-Guided Diffusion Reasoner for Neuro-Symbolic Learning (AAAI 2026 · similarity 0.87)
- LogicMP: A Neuro-symbolic Approach for Encoding First-order Logic Constraints (ICLR 2024 · similarity 0.86)
- Training Binary Neural Networks via Gaussian Variational Inference and Low-Rank Semidefinite Programming (NeurIPS 2024 · similarity 0.85)
- Taming Binarized Neural Networks and Mixed-Integer Programs (AAAI 2024 · similarity 0.85)
- DeepSaDe: Learning Neural Networks That Guarantee Domain Constraint Satisfaction (AAAI 2024 · similarity 0.85)
- Enforcing Hard Linear Constraints in Deep Learning Models with Decision Rules (NeurIPS 2025 · 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.
