Research map / Learning theory and algorithms / Combinatorial and search optimization
Neural Fields with Hard Constraints of Arbitrary Differential Order
Fangcheng Zhong, Kyle Thomas Fogarty, Param Hanji, Tianhao Walter Wu, Alejandro Sztrajman, Andrew Everett Spielberg, Andrea Tagliasacchi, Petra Bosilj, Cengiz Oztireli
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 exact normal constraints instead of pseudo-normal constraints (ablation) | CNF hypernet kernel · 2D circle normal interpolation | mean normal error: 0.499 -> 1.083\times 10^{-6} | Improved |
| use exact normal constraints instead of pseudo-normal constraints (ablation) | CNF hypernet kernel · 2D triangle normal interpolation | mean normal error: 0.238 -> 6.292\times 10^{-6} | Improved |
| use exact normal constraints instead of SIREN | CNF · 2D diamond normal interpolation | mean normal error: 1.05 \times 10^{-1} -> 3.368\times 10^{-6} | Improved |
| train after initialization with Eikonal loss | CNF exact normal · 2D line normal interpolation | mean normal error: 2.737\times 10^{-6} -> 6.963\times 10^{-6} | Hurt |
| pre-train skewed RBF hyperparameters on mu=0 instead of random init (ablation) | Skewed RBF · 1D advection, shift mu=10000 | RMSE: 635.78 -> 23.844 | Improved |
| pre-train skewed RBF hyperparameters instead of random init (ablation) | Skewed RBF · 1D advection, no shift | RMSE: 0.0127 -> 0.0070 | Improved |
| pre-train skewed RBF hyperparameters instead of random init (ablation) | Skewed RBF · 1D advection, shift mu=10 | nRMSE: 0.0594 -> 0.0018 | Improved |
| replace Gaussian kernel basis with dot-product kernel (ablation) | CNF with SIREN encoder · BRDF fitting (MERL) | reconstruction quality | Hurt |
| replace Gaussian kernel basis with hypernet basis or constraint layer (ablation) | CNF · BRDF fitting (MERL) | reconstruction quality | Hurt |
| use Gaussian kernel basis instead of other neural basis functions (ablation) | CNF · 2D function with 4096 constraints | condition number | Improved |
| use skewed RBF with learnable per-basis covariance instead of fixed RBF (ablation) | Kansa-style CNF · 1D advection on irregular grid | error | Improved |
| pretrain encoder weights as circle SDF | CNF hypernet kernel · 2D shape reconstruction | plausibility of surface | Improved |
Closest papers by meaning
- Enforcing Hard Linear Constraints in Deep Learning Models with Decision Rules (NeurIPS 2025 · similarity 0.86)
- DeepSaDe: Learning Neural Networks That Guarantee Domain Constraint Satisfaction (AAAI 2024 · similarity 0.85)
- Fast Training of Sinusoidal Neural Fields via Scaling Initialization (ICLR 2025 · similarity 0.85)
- How to Train Neural Field Representations: A Comprehensive Study and Benchmark (CVPR 2024 · similarity 0.84)
- On the Hardness of Training Deep Neural Networks Discretely (AAAI 2025 · similarity 0.84)
- Scaling physics-informed hard constraints with mixture-of-experts (ICLR 2024 · 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.
